This resource is using open-source code maintained in github (see the quick-start-guide section) and available for download from NGC

The Deep Learning Recommendation Model (DLRM) is a recommendation model designed to make use of both categorical and numerical inputs.\nIt was first described in Deep Learning Recommendation Model for Personalization and Recommendation Systems.\nThis repository provides a reimplementation of the code-base provided originally here.\nThe scripts enable you to train DLRM on the Criteo Terabyte Dataset.

Using the scripts provided here, you can efficiently train models that are too large to fit into a single GPU.\nThis is because we use a hybrid-parallel approach, which combines model parallelism with data parallelism for\ndifferent parts of the neural network.\nThis is explained in details in the next section.

This model uses a slightly different preprocessing procedure than the one found in the original implementation.\nMost importantly, we use a technique called frequency thresholding to demonstrate models of different size.\nThe smallest model can be trained on a single V100-32GB GPU, while the largest one needs 8xA100-80GB GPUs.
\nThe table below summarizes the model sizes and frequency thresholds used in this repository:

\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n

tr:nth-child(odd)]:bg-surface-sunken\">

Name	Frequency threshold	Number of parameters	Model size
Small	15	4.2B	15.6 GiB
Large	3	22.8B	84.9 GiB
Extra large	0	113B	421 GiB

You can find a detailed description of the preprocessing steps in the Dataset guidelines section.

Using DLRM, you can train a high-quality general model for recommendations.

This model is trained with mixed precision using Tensor Cores on Volta, Turing and NVIDIA Ampere GPU architectures.\nTherefore, researchers can get results 2x faster than training without Tensor Cores while experiencing the\nbenefits of mixed precision training. This model is tested against each NGC monthly container\nrelease to ensure consistent accuracy and performance over time.

Model architecture

DLRM accepts two types of features: categorical and numerical. For each categorical feature, an embedding table is used to provide dense representation to each unique value. The dense features enter the model and are transformed by a simple neural network referred to as \"bottom MLP\".

This part of the network consists of a series\nof linear layers with ReLU activations. The output of the bottom MLP and the embedding vectors are then fed into the \"dot interaction\" operation. The output of \"dot interaction\" is then concatenated with the features resulting from bottom MLP and fed into the \"top MLP\" which is a series of dense layers with activations.\nThe model outputs a single number which can be interpreted as a likelihood of a certain user clicking an ad.

\n \n
\nFigure 1. The architecture of DLRM.\n

Default configuration

The following features were implemented in this model:

general\n
- static loss scaling for Tensor Cores (mixed precision) training
- hybrid-parallel multi-GPU training
\n
preprocessing\n
- dataset preprocessing using Spark 3 on GPUs
\n

Feature support matrix

The following features are supported by this model:

\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n\n

tr:nth-child(odd)]:bg-surface-sunken\">

Feature	DLRM
Automatic mixed precision (AMP)	Yes
XLA	Yes
Hybrid-parallel training with Merlin Distributed Embeddings	Yes
Preprocessing on GPU with Spark 3	Yes
Multi-node training	Yes

Features

Automatic Mixed Precision (AMP)\nEnables mixed precision training without any changes to the code-base by performing automatic graph rewrites and loss scaling controlled by an environmental variable.

XLA

The training script supports a --xla flag. It can be used to enable XLA JIT compilation. Currently, we use XLA Lite. It delivers a steady 10-30% performance boost depending on your hardware platform, precision, and the number of GPUs. It is turned off by default.

Horovod\nHorovod is a distributed training framework for TensorFlow, Keras, PyTorch, and MXNet. The goal of Horovod is to make distributed deep learning fast and easy to use. For more information about how to get started with Horovod, see the Horovod official repository.

Hybrid-parallel training with Merlin Distributed Embeddings\nOur model uses Merlin Distributed Embeddings to implement efficient multi-GPU training.\nFor details, see example sources in this repository or see the TensorFlow tutorial.\nFor the detailed description of our multi-GPU approach, visit this section.

Multi-node training\nThis repository supports multinode training. For more information refer to the multinode section

Mixed precision training

Mixed precision is the combined use of different numerical precisions in a computational method. Mixed precision training offers significant computational speedup by performing operations in half-precision format while storing minimal information in single-precision to retain as much information as possible in critical parts of the network. Since the introduction of Tensor Cores in Volta, and following with both the Turing and Ampere architectures, significant training speedups are experienced by switching to mixed precision -- up to 3.4x overall speedup on the most arithmetically intense model architectures. Using mixed precision training requires two steps:

Porting the model to use the FP16 data type where appropriate.
Adding loss scaling to preserve small gradient values.

The ability to train deep learning networks with lower precision was introduced in the Pascal architecture and first supported in CUDA 8 in the NVIDIA Deep Learning SDK.

For information about:

How to train using mixed precision, see the Mixed Precision Training paper and Training With Mixed Precision documentation.
Techniques used for mixed precision training, see the Mixed-Precision Training of Deep Neural Networks blog.

Enabling mixed precision

Mixed precision training is turned off by default. To turn it on, issue the --amp flag to the main.py script.

Enabling TF32

TensorFloat-32 (TF32) is the new math mode in NVIDIA A100 GPUs for handling the matrix math also called tensor operations. TF32 running on Tensor Cores in A100 GPUs can provide up to 10x speedups compared to single-precision floating-point math (FP32) on Volta GPUs.

TF32 Tensor Cores can speed up networks using FP32, typically with no loss of accuracy. It is more robust than FP16 for models which require high dynamic range for weights or activations.

For more information, refer to the TensorFloat-32 in the A100 GPU Accelerates AI Training, HPC up to 20x blog post.

TF32 is supported in the NVIDIA Ampere GPU architecture and is enabled by default.

Hybrid-parallel training with Merlin Distributed Embeddings

Many recommendation models contain very large embedding tables. As a result, the model is often too large to fit onto a single device.\nThis could be easily solved by training in a model-parallel way, using either the CPU or other GPUs as \"memory donors\".\nHowever, this approach is suboptimal as the \"memory donor\" devices' compute is not utilized.\nIn this repository, we use the model-parallel approach for the Embedding Tables while employing a usual data parallel approach\nfor the more compute-intensive MLPs and Dot Interaction layer. This way, we can train models much larger than what would normally fit into\na single GPU while at the same time making the training faster by using multiple GPUs. We call this approach hybrid-parallel training.

To implement this approach we use the Merlin Distributed Embeddings library.\nIt provides a scalable model parallel wrapper called distributed_embeddings.dist_model_parallel. This wrapper automatically distributes embedding tables to multiple GPUs.\nThis way embeddings can be scaled beyond single GPU's memory capacity without\ncomplex code to handle cross-worker communication.

Under the hood, Merlin Distributed Embeddings uses a\nspecific multi-GPU communication pattern called\nall-2-all to transition from model-parallel to data-parallel\nparadigm. In the original DLRM whitepaper this has been referred to as \"butterfly shuffle\".

An example model using Hybrid Parallelism is shown in Figure 2. The compute intensive dense layers are run in data-parallel\nmode. The smaller embedding tables are run model-parallel, such that each smaller table is placed entirely on a single device.\nThis is not suitable for larger tables that need more memory than can be provided by a single device. Therefore,\nthose large tables are split into multiple parts and each part is run on a different GPU.

\n \n
\nFigure 2. Hybrid parallelism with Merlin Distributed Embeddings.\n

In this repository we train models of three sizes: \"small\" (15.6 GiB), \"large\" (84.9 GiB) and \"extra large\" (421 GiB).\nThe \"small\" model can be trained on a single V100-32GB GPU. The \"large\" model needs at least 8xV100-32GB GPUs,\nbut each of the tables it uses can fit on a singleGPU.

The \"extra large\" model, on the other hand, contains tables that do not fit into a singledevice, and will be automatically\nsplit and stored across multiple GPUs by Merlin Distributed Embeddings.

Training very large embedding tables

We tested this approach by training a DLRM model on the Criteo Terabyte dataset with the frequency limiting option turned off (set to zero).\nThe weights of the resulting model take 421 GiB. The largest table weighs 140 GiB.\nHere are the commands you can use to reproduce this:

# build and run the preprocessing container as in the Quick Start Guide\n# then when preprocessing set the frequency limit to 0:\n./prepare_dataset.sh DGX2 0\n\n# build and run the training container same as in the Quick Start Guide\n# then append options necessary for training very large embedding tables:\nhorovodrun -np 8 -H localhost:8 --mpi-args=--oversubscribe numactl --interleave=all -- python -u main.py --dataset_path /data/dlrm/ --amp --xla\n

When using this method on a DGX A100 with 8 A100-80GB GPUs and a large-enough dataset, it is possible to train a single embedding table of up to 600 GB. You can also use multi-node training (described below) to train even larger recommender systems.

This mode was used to train the 421GiB \"extra large\" model in the DGX A100-80G performance section.

Multi-node training

Multi-node training is supported. Depending on the exact interconnect hardware and model configuration,\nyou might experience only a modest speedup with multi-node.\nMulti-node training can also be used to train larger models.\nFor example, to train a 1.68 TB variant of DLRM on multi-node, you can run:

cmd='numactl --interleave=all -- python -u main.py --dataset_path /data/dlrm/full_criteo_data --amp --xla\\\n--embedding_dim 512 --bottom_mlp_dims 512,256,512' \\\nsrun_flags='--mpi=pmix' \\\ncont=nvidia_dlrm_tf \\\nmounts=/data/dlrm:/data/dlrm \\\nsbatch -n 32 -N 4 -t 00:20:00 slurm_multinode.sh\n

Preprocessing on GPU with Spark 3

Refer to the \"Preprocessing with Spark\" section for a detailed description of the Spark 3 GPU functionality.

BYO dataset functionality overview

This section describes how you can train the DeepLearningExamples RecSys models on your own datasets without changing\nthe model or data loader and with similar performance to the one published in each repository.\nThis can be achieved thanks to Dataset Feature Specification, which describes how the dataset, data loader and model\ninteract with each other during training, inference and evaluation.\nDataset Feature Specification has a consistent format across all recommendation models in NVIDIA's DeepLearningExamples\nrepository, regardless of dataset file type and the data loader,\ngiving you the flexibility to train RecSys models on your own datasets.

Glossary

The Dataset Feature Specification consists of three mandatory and one optional section:

feature_spec provides a base of features that may be referenced in other sections, along with their metadata.\nFormat: dictionary (feature name) => (metadata name => metadata value)

source_spec provides information necessary to extract features from the files that store them.\nFormat: dictionary (mapping name) => (list of chunks)

Mappings are used to represent different versions of the dataset (think: train/validation/test, k-fold splits). A mapping is a list of chunks.
Chunks are subsets of features that are grouped together for saving. For example, some formats may constrain data saved in one file to a single data type. In that case, each data type would correspond to at least one chunk. Another example where this might be used is to reduce file size and enable more parallel loading. Chunk description is a dictionary of three keys:
\n
- type provides information about the format in which the data is stored. Not all formats are supported by all models.
- features is a list of features that are saved in a given chunk. Order of this list may matter: for some formats, it is crucial for assigning read data to the proper feature.
- files is a list of paths to files where the data is saved. For Feature Specification in yaml format, these paths are assumed to be relative to the yaml file's directory (basename). Order of this list matters: It is assumed that rows 1 to i appear in the first file, rows i+1 to j in the next one, etc.
\n

channel_spec determines how features are used. It is a mapping (channel name) => (list of feature names).

Channels are model specific magic constants. In general, data within a channel is processed using the same logic. Example channels: model output (labels), categorical ids, numerical inputs, user data, and item data.

metadata is a catch-all, wildcard section: If there is some information about the saved dataset that does not fit into the other sections, you can store it here.

Dataset feature specification

Data flow can be described abstractly:\nInput data consists of a list of rows. Each row has the same number of columns; each column represents a feature.\nThe columns are retrieved from the input files, loaded, aggregated into channels and supplied to the model/training script.

FeatureSpec contains metadata to configure this process and can be divided into three parts:

\n
Specification of how data is organized on disk (source_spec). It describes which feature (from feature_spec) is stored in which file and how files are organized on disk.
\n
\n
Specification of features (feature_spec). Describes a dictionary of features, where key is feature name and values are features' characteristics such as dtype and other metadata (for example, cardinalities for categorical features)
\n
\n
Specification of model's inputs and outputs (channel_spec). Describes a dictionary of model's inputs where keys specify model channel's names and values specify lists of features to be loaded into that channel. Model's channels are groups of data streams to which common model logic is applied, for example categorical/continuous data, user/item ids. Required/available channels depend on the model
\n

The FeatureSpec is a common form of description regardless of underlying dataset format, dataset data loader form and model.

Data flow in NVIDIA Deep Learning Examples recommendation models

The typical data flow is as follows:

S.0. Original dataset is downloaded to a specific folder.
S.1. Original dataset is preprocessed into Intermediary Format. For each model, the preprocessing is done differently, using different tools. The Intermediary Format also varies (for example, for DLRM PyTorch, the Intermediary Format is a custom binary one.)
S.2. The Preprocessing Step outputs Intermediary Format with dataset split into training and validation/testing parts along with the Dataset Feature Specification yaml file. Metadata in the preprocessing step is automatically calculated.
S.3. Intermediary Format data together with Dataset Feature Specification are fed into training/evaluation scripts. Data loader reads Intermediary Format and feeds the data into the model according to the description in the Dataset Feature Specification.
S.4. The model is trained and evaluated

\n \n
\n

Fig.1. Data flow in Recommender models in NVIDIA Deep Learning Examples repository. Channels of the model are drawn in green.

Example of dataset feature specification

As an example, let's consider a Dataset Feature Specification for a small CSV dataset for some abstract model.

feature_spec:\n  user_gender:\n    dtype: torch.int8\n    cardinality: 3 #M,F,Other\n  user_age: #treated as numeric value\n    dtype: torch.int8\n  user_id:\n    dtype: torch.int32\n    cardinality: 2655\n  item_id:\n    dtype: torch.int32\n    cardinality: 856\n  label:\n    dtype: torch.float32\n\nsource_spec:\n  train:\n    - type: csv\n      features:\n        - user_gender\n        - user_age\n      files:\n        - train_data_0_0.csv\n        - train_data_0_1.csv\n    - type: csv\n      features:\n        - user_id\n        - item_id\n        - label\n      files:\n        - train_data_1.csv\n  test:\n    - type: csv\n      features:\n        - user_id\n        - item_id\n        - label\n        - user_gender\n        - user_age\n        \n      files:\n        - test_data.csv\n\nchannel_spec:\n  numeric_inputs: \n    - user_age\n  categorical_user_inputs: \n    - user_gender\n    - user_id\n  categorical_item_inputs: \n    - item_id\n  label_ch: \n    - label\n

The data contains five features: (user_gender, user_age, user_id, item_id, label). Their data types and necessary metadata are described in the feature specification section.

In the source mapping section, two mappings are provided: one describes the layout of the training data, the other of the testing data. The layout for training data has been chosen arbitrarily to showcase the flexibility.\nThe train mapping consists of two chunks. The first one contains user_gender and user_age, saved as a CSV, and is further broken down into two files. For specifics of the layout, refer to the following example and consult the glossary. The second chunk contains the remaining columns and is saved in a single file. Notice that the order of columns is different in the second chunk - this is alright, as long as the order matches the order in that file (that is, columns in the .csv are also switched)

Let's break down the train source mapping. The table contains example data color-paired to the files containing it.

\n\n

The channel spec describes how the data will be consumed. Four streams will be produced and available to the script/model.\nThe feature specification does not specify what happens further: names of these streams are only lookup constants defined by the model/script.\nBased on this example, we can speculate that the model has three input channels: numeric_inputs, categorical_user_inputs,\ncategorical_item_inputs, and one output channel: label.\nFeature names are internal to the FeatureSpec and can be freely modified.

BYO dataset functionality

In order to train any Recommendation model in NVIDIA Deep Learning Examples one can follow one of three possible ways:

\n
One delivers already preprocessed dataset in the Intermediary Format supported by data loader used by the training script\n(different models use different data loaders) together with FeatureSpec yaml file describing at least specification of dataset, features and model channels
\n
\n
One uses a transcoding script
\n
\n
One delivers dataset in non-preprocessed form and uses preprocessing scripts that are a part of the model repository.\nIn order to use already existing preprocessing scripts, the format of the dataset needs to match the one of the original datasets.\nThis way, the FeatureSpec file will be generated automatically, but the user will have the same preprocessing as in the original model repository.
\n

","resourceVersions",[230,231],"resourceVersionsTotal","hasMoreVersions","versionMetadata",{"_209":28,"_210":-5,"_52":117,"_114":-5,"_118":211,"_120":-5,"_212":116,"_125":16,"_122":34,"_213":-5,"_121":34,"_214":215,"_216":91,"_217":218,"_219":91,"_220":221,"_222":91,"_223":224,"_225":91,"_226":227,"_228":91,"_123":229},"versionNotFound","isLoggedIn","latestVersionPath","/orgs/nvidia/dle/resources/dlrm_tf2/-","imageReference","cliDownloadCommand","ngc registry resource download-version nvidia/dle/dlrm_tf2:22.06.0","helmFetchCommand","viewAllVersionsPath","/orgs/nvidia/dle/resources/dlrm_tf2/22.06.0/version-history","accessCheckResult",{"_208":91},"hasRelatedCollections","context",{"_204":205,"_206":207},1787557791872,[],[],[198,199,200,201,202,203],{"_48":63,"_50":74,"_52":65,"_54":66,"_56":75,"_58":59},{"_48":63,"_50":70,"_52":65,"_54":66,"_56":71,"_58":59},{"_48":63,"_50":68,"_52":65,"_54":66,"_56":69,"_58":59},{"_48":63,"_50":64,"_52":65,"_54":66,"_56":67,"_58":59},{"_48":60,"_50":61,"_52":147,"_54":-5,"_56":-5,"_58":59},{"_48":49,"_50":51,"_52":53,"_54":55,"_56":57,"_58":59},"orgName","nvidia","teamName","dle","allowed","versionId","description",["D",1663857111272],"totalSizeInBytes","malwareScanStatus","releaseNotes","We're constantly refining and improving our performance on AI and HPC workloads even on the same hardware with frequent updates to our software stack. For our latest performance data please refer to these pages for [AI](https://developer.nvidia.com/deep-learning-performance-training-inference) and [HPC](https://developer.nvidia.com/hpc-application-performance) benchmarks.\n\n### Changelog\nJuly 2022\n- Start using Merlin Distributed Embeddings\n\nMarch 2022\n- Major performance improvements\n- Support for BYO dataset\n\nMarch 2021\n- Initial release\n\n### Known issues\n\n#### Checkpointing\nTensorFlow runs into issues when trying to save model checkpoints for extremely large variables.\nWe circumvent this by using a custom checkpoint format that splits the variables into pieces and stores each piece independently.\nHowever, this custom format cannot be used by the standard inference deployment frameworks such as ONNX.\n\n#### Inference performance\nCurrent inference performance was evaluated in python using TensorFlow 2.9.1.\nThis provides ease of use and flexibility but is suboptimal in terms of performance.\nIf you're interested in state-of-the-art performance for recommender system inference,\nplease review our results in [the MLPerf v0.7 benchmark](https://mlperf.org/inference-results/)\nwhere we used [TensorRT](https://developer.nvidia.com/tensorrt).\nYou might also want to check [the source code of our MLPerf Inference submission](https://github.com/mlcommons/inference_results_v0.7/tree/master/closed/NVIDIA/code/dlrm/tensorrt).","hasReleaseNotes","setup","The following section lists the requirements for training DLRM.\n\n### Requirements\n\nThis repository contains Dockerfile which extends the TensorFlow 2 NGC container and encapsulates some dependencies. Aside from these dependencies, ensure you have the following components:\n- [NVIDIA Docker](https://github.com/NVIDIA/nvidia-docker)\n- [TensorFlow 2 21.02-py3](https://ngc.nvidia.com/catalog/containers/nvidia:tensorflow/tags) NGC container\n- Supported GPUs:\n - [NVIDIA Volta architecture](https://www.nvidia.com/en-us/data-center/volta-gpu-architecture/)\n - [NVIDIA Turing architecture](https://www.nvidia.com/en-us/geforce/turing/)\n - [NVIDIA Ampere architecture](https://www.nvidia.com/en-us/data-center/nvidia-ampere-gpu-architecture/)\n\nFor more information about how to get started with NGC containers, see the following sections from the NVIDIA GPU Cloud Documentation and the Deep Learning Documentation:\n- [Getting Started Using NVIDIA GPU Cloud](https://docs.nvidia.com/ngc/ngc-getting-started-guide/index.html)\n- [Accessing And Pulling From The NGC Container Registry](https://docs.nvidia.com/deeplearning/frameworks/user-guide/index.html#accessing_registry)\n- [Running TensorFlow](https://docs.nvidia.com/deeplearning/frameworks/tensorflow-release-notes/running.html#running)\n\nFor those unable to use the TensorFlow NGC container, to set up the required environment or create your own container, see the versioned [NVIDIA Container Support Matrix](https://docs.nvidia.com/deeplearning/frameworks/support-matrix/index.html).","hasSetup","quickStartGuide","To train your model using mixed or TF32 precision with Tensor Cores or using FP32, perform the following steps using\nthe default parameters of DLRM on the Criteo Terabyte dataset. For the specifics concerning training and inference,\nsee the [Advanced](#advanced) section.\n\n1. Clone the repository.\n```\ngit clone https://github.com/NVIDIA/DeepLearningExamples\ncd DeepLearningExamples/TensorFlow2/Recommendation/DLRM\n```\n\n2. Build a DLRM Docker container.\n```bash\ndocker build -t nvidia_dlrm_tf .\ndocker build -t nvidia_dlrm_spark -f Dockerfile_spark .\n```\n\n3. Start an interactive session in the NGC container to run preprocessing.\nThe DLRM TensorFlow container can be launched with:\n```bash\nmkdir -p data\ndocker run --runtime=nvidia -it --rm --ipc=host -v ${PWD}/data:/data nvidia_dlrm_spark bash\n```\n\n4. Download and preprocess the dataset.\n\nYou can download the data by following the instructions at: http://labs.criteo.com/2013/12/download-terabyte-click-logs/.\n\nWhen you have successfully downloaded the dataset, put it in the `/data/dlrm/criteo/` directory in the container (`$PWD/data/dlrm/criteo` in the host system).\n\nHere are a few examples of different preprocessing commands. For the details on how those scripts work and detailed description of all the parameters, consult the [preprocess with spark section](#preprocess-with-spark).\n\n```bash\ncd preproc\n\n# to run on a DGX-2 with a frequency limit of 3 (will need 8xV100-32GB to fit the model in GPU memory)\n./prepare_dataset.sh DGX2 3\n\n# to run on a DGX-2 with a frequency limit of 15 (should fit on a single V100-32GB):\n./prepare_dataset.sh DGX2 15\n#\n# to run on CPU with a frequency limit of 15:\n./prepare_dataset.sh CPU 15\n\n# to run on DGX-2 with no frequency limit:\n./prepare_dataset.sh DGX2 0\n```\n\n5. Start training.\n\nFirst, start the Docker container:\n```bash\ndocker run --cap-add SYS_NICE --runtime=nvidia -it --rm --ipc=host -v ${PWD}/data:/data nvidia_dlrm_tf bash\n```\n\n- single-GPU A100-80GB:\n```bash\nhorovodrun -np 1 -H localhost:1 --mpi-args=--oversubscribe numactl --interleave=all -- python -u main.py --dataset_path /data/dlrm/ --amp --xla --save_checkpoint_path /data/dlrm/checkpoint/dlrm\n```\n\n- single-GPU V100-32GB:\n```bash\nhorovodrun -np 1 -H localhost:1 --mpi-args=--oversubscribe numactl --interleave=all -- python -u main.py --dataset_path /data/dlrm/ --xla --save_checkpoint_path /data/dlrm/checkpoint/dlrm\n```\n\n- multi-GPU for DGX A100 (model size 90GiB or 421GiB depending on the dataset passed)\n```bash\nhorovodrun -np 8 -H localhost:8 --mpi-args=--oversubscribe numactl --interleave=all -- python -u main.py --dataset_path /data/dlrm/ --amp --xla --save_checkpoint_path /data/dlrm/checkpoint/dlrm\n```\n\n- multi-GPU for DGX2 (model size 90GiB):\n```bash\nhorovodrun -np 16 -H localhost:16 --mpi-args=--oversubscribe numactl --interleave=all -- python -u main.py --dataset_path /data/dlrm/ --amp --xla --column_slice_threshold 5000000000 --save_checkpoint_path /data/dlrm/checkpoint/dlrm \n```\n\n- multi-GPU for DGX1V-32GB (model size 90GiB):\n```bash\nhorovodrun -np 8 -H localhost:8 --mpi-args=--oversubscribe numactl --interleave=all -- python -u main.py --dataset_path /data/dlrm/ --amp --xla --column_slice_threshold 5000000000 --save_checkpoint_path /data/dlrm/checkpoint/dlrm\n```\n\n6. Start evaluation.\n\nTo evaluate a previously trained checkpoint, append `--restore_checkpoint_path --mode eval` to the command used for training. For example, to test a checkpoint trained on 8xA100 80GB, run:\n\n```bash\nhorovodrun -np 8 -H localhost:8 --mpi-args=--oversubscribe numactl --interleave=all -- python -u main.py --dataset_path /data/dlrm/ --amp --xla --restore_checkpoint_path /data/dlrm/checkpoint/dlrm --mode eval\n```","hasQuickStartGuide","advanced","The following sections provide greater details of the dataset, running training and inference, and the training results.\n\n### Scripts and sample code\n\nThese are the important modules in this repository:\n- `main.py` - The main entrypoint script for training, evaluating, and benchmarking.\n- `model.py` - Defines the DLRM model and some auxiliary functions used to train it.\n- `dataloader.py` - Handles defining the dataset objects based on command-line flags.\n- `datasets.py` - Defines the `TfRawBinaryDataset` class responsible for storing and loading the training data.\n- `slurm_multinode.sh` - Example batch script for multi-node training on SLURM clusters.\n- `lr_scheduler.py` - Defines a TensorFlow learning rate scheduler that supports both learning rate warmup and polynomial decay.\n- `embedding.py` - Implementations of the embedding layers.\n- `interaction.py` - Implementation of the dot-interaction layer using TensorFlow operations.\n- `tensorflow-dot-based-interact` - A directory with a set of custom CUDA kernels. They provide fast implementations of the dot-interaction operation for various precisions and hardware platforms.\n- `utils.py` - General utilities, such as a timer used for taking performance measurements.\n\n### Parameters\n\nThe table below lists the most important command-line parameters of the `main.py` script.\n\n| Scope| parameter| Comment| Default Value |\n| ----- | --- | ---- | ---- |\n|datasets|dataset_path|Path to the JSON file with the sizes of embedding tables|\n|function|mode| Choose \"train\" to train the model, \"inference\" to benchmark inference and \"eval\" to run validation| train|\n|optimizations|amp| Enable automatic mixed precision| False\n|optimizations|xla| Enable XLA| False|\n|hyperparameters|batch_size| Batch size used for training|65536|\n|hyperparameters|epochs| Number of epochs to train for|1|\n|hyperparameters|optimizer| Optimization algorithm for training |SGD|\n|hyperparameters|evals_per_epoch| Number of evaluations per epoch|1|\n|hyperparameters|valid_batch_size| Batch size used for validation|65536|\n|hyperparameters|max_steps| Stop the training/inference after this many optimization steps|-1|\n|checkpointing|restore_checkpoint_path| Path from which to restore a checkpoint before training|None|\n|checkpointing|save_checkpoint_path| Path to which to save a checkpoint file at the end of the training|None|\n|debugging|run_eagerly| Disable all tf.function decorators for debugging|False|\n|debugging|print_freq| Number of steps between debug prints|1000|\n\n### Command-line options\n\nThe `main.py` script supports a number of command-line flags. You can get the descriptions of those by running `python main.py --help`.\n\n### Getting the data\n\nThis example uses the [Criteo Terabyte Dataset](https://labs.criteo.com/2013/12/download-terabyte-click-logs/).\nThe first 23 days are used as the training set. The last day is split in half. The first part is used as a validation set and the second set is used as a hold-out test set.\n\n#### Dataset guidelines\n\nThe preprocessing steps applied to the raw data include:\n- Replacing the missing values with `0`.\n- Replacing the categorical values that exist fewer than 15 times with a special value.\n- Converting the hash values to consecutive integers.\n- Adding 2 to all the numerical features so that all of them are greater or equal to 1.\n- Taking a natural logarithm of all numerical features.\n\n#### BYO dataset \n\nThis implementation supports using other datasets thanks to BYO dataset functionality. \nThe BYO dataset functionality allows users to plug in their dataset in a common fashion for all Recommender models \nthat support this functionality. Using BYO dataset functionality, the user does not have to modify the source code of \nthe model thanks to the Feature Specification file. For general information on how BYO dataset works, refer to the \n[BYO dataset overview section](#byo-dataset-functionality-overview).\n\nThere are three ways to plug in user's dataset:\n

1. Provide an unprocessed dataset in a format matching the one used by Criteo 1TB, then use Criteo 1TB's preprocessing. Feature Specification file is then generated automatically.

\nThe required format of the user's dataset is:\n\nThe data should be split into text files. Each line of those text files should contain a single training example. \nAn example should consist of multiple fields separated by tabulators:\n\n* The first field is the label – 1 for a positive example and 0 for negative.\n* The next N tokens should contain the numerical features separated by tabs.\n* The next M tokens should contain the hashed categorical features separated by tabs.\n\nThe correct dataset files together with the Feature Specification yaml file will be generated automatically by preprocessing script.\n\nFor an example of using this process, refer to the [Quick Start Guide](#quick-start-guide)\n\n

\n\n

2. Provide a CSV containing preprocessed data and a simplified Feature Specification yaml file, then transcode the data with `transcode.py` script

\nThis option should be used if the user has their own CSV file with a preprocessed dataset they want to train on.\n\nThe required format of the user's dataset is:\n* CSV files containing the data, already split into train and test sets. \n* Feature Specification yaml file describing the layout of the CSV data\n\nFor an example of a feature specification file, refer to the `tests/transcoding` folder.\n\nThe CSV containing the data:\n* should be already split into train and test\n* should contain no header\n* should contain one column per feature, in the order specified by the list of features for that chunk \n in the source_spec section of the feature specification file\n* categorical features should be non-negative integers in the range [0,cardinality-1] if cardinality is specified\n\nThe Feature Specification yaml file:\n* needs to describe the layout of data in CSV files\n* may contain information about cardinalities. However, if set to `auto`, they will be inferred from the data by the transcoding script.\n\nRefer to `tests/transcoding/small_csv.yaml` for an example of the yaml Feature Specification.\n\nThe following example shows how to use this way of plugging user's dataset:\n\nPrepare your data and save the path:\n```bash\nDATASET_PARENT_DIRECTORY=/raid/dlrm\n```\n\nBuild the DLRM image with:\n```bash\ndocker build -t nvidia_dlrm_tf .\n```\nLaunch the container with:\n```bash\ndocker run --cap-add SYS_NICE --runtime=nvidia -it --rm --ipc=host -v ${DATASET_PARENT_DIRECTORY}/data:/data nvidia_dlrm_tf bash\n```\n\nIf you are just testing the process, you can create synthetic csv data:\n```bash\npython gen_csv.py --feature_spec_in tests/transcoding/small_csv.yaml\n```\n\nConvert the data:\n```bash\nmkdir /data/conversion_output\ncp tests/transcoding/small_csv.yaml /data/feature_spec.yaml\npython transcode.py --input /data --output /data/converted\n```\nYou may need to tune the --chunk_size parameter. Higher values speed up the conversion but require more RAM.\n\nThis will convert the data from `/data` and save the output in `/data/converted`.\nA feature specification file describing the new data will be automatically generated.\n\nTo run the training on 1 GPU:\n```bash\nhorovodrun -np 1 -H localhost:1 --mpi-args=--oversubscribe numactl --interleave=all -- python -u main.py --dataset_path /data/converted --amp --xla\n```\n\n- multi-GPU for DGX A100:\n```bash\nhorovodrun -np 8 -H localhost:8 --mpi-args=--oversubscribe numactl --interleave=all -- python -u main.py --dataset_path /data/converted --amp --xla\n```\n\n- multi-GPU for DGX-1 and DGX-2:\n```bash\nhorovodrun -np 8 -H localhost:8 --mpi-args=--oversubscribe numactl --interleave=all -- python -u main.py --dataset_path /data/converted --amp --xla\n```\n

3. Provide a fully preprocessed dataset, saved in split binary files, and a Feature Specification yaml file

\nThis is the option to choose if you want full control over preprocessing and/or want to preprocess data directly to the target format.\n\nYour final output will need to contain a Feature Specification yaml describing data and file layout. \nFor an example feature specification file, refer to `tests/feature_specs/criteo_f15.yaml`\n\nFor details, refer to the [BYO dataset overview section](#byo-dataset-functionality-overview).\n

\n\n##### Channel definitions and requirements\n\nThis model defines three channels:\n\n- categorical, accepting an arbitrary number of features\n- numerical, accepting an arbitrary number of features\n- label, accepting a single feature\n\nThe training script expects two mappings:\n\n- train\n- test\n\nFor performance reasons:\n* The only supported dataset type is split binary\n* Splitting chunks into multiple files is not supported.\n* Each categorical feature has to be provided in a separate chunk\n* All numerical features have to be provided in a single chunk\n* All numerical features have to appear in the same order in channel_spec and source_spec\n* Only integer types are supported for categorical features\n* Only float16 is supported for numerical features\n\n##### BYO dataset constraints for the model\n\nThere are the following constraints of BYO dataset functionality for this model:\n1. The performance of the model depends on the dataset size. Generally, the model should scale better for datasets containing more data points. For a smaller dataset, you might experience slower performance than the one reported for Criteo\n2. Using other datasets might require tuning some hyperparameters (for example, learning rate, beta1 and beta2) to reach desired accuracy.\n3. The optimized cuda interaction kernels for FP16 and TF32 assume that the number of categorical variables is smaller than WARP_SIZE=32 and embedding size is <=128\n\n#### Preprocess with Spark\n\nThe preprocessing scripts provided in this repository support running both on CPU and on DGX-2 using [Apache Spark 3.0](https://www.nvidia.com/en-us/deep-learning-ai/solutions/data-science/apache-spark-3/).\nIt should be possible to change the values in `preproc/dgx2_config.sh`\nso that they'll work on other hardware platforms such as DGX-1.\n\nNote that the preprocessing will require about 4TB of disk storage.\n\nThe syntax for the preprocessing script is as follows:\n```bash\ncd preproc\n./prepare_dataset.sh \n```\n\nThe first argument is the hardware platform to use (either DGX-2 or pure-CPU). The second argument means the frequency\nthreshold to apply to the categorical variables. For a frequency threshold `T`, the categorical values that occur less\noften than `T` will be replaced with a special embedding. Thus, a larger value of `T` will require smaller embedding tables\nand will substantially reduce the overall size of the model.\n\nFor the Criteo Terabyte dataset we recommend a frequency threshold of `T=3` if you intend to run the hybrid-parallel mode\non multiple GPUs. If you want to make the model fit into a single NVIDIA Tesla V100-32GB, you can set `T=15`.\n\nThe preprocessing scripts makes use of the following environment variables to configure the data directory paths:\n- `download_dir` – this directory should contain the original Criteo Terabyte CSV files\n- `spark_output_path` – directory to which the parquet data will be written\n- `conversion_intermediate_dir` – directory used for storing intermediate data used to convert from parquet to train-ready format\n- `final_output_dir` – directory to store the final results of the preprocessing which can then be used to train DLRM\n\nThe script `spark_data_utils.py` is a PySpark application, which is used to preprocess the Criteo Terabyte Dataset. In the Docker image, we have installed Spark 3.0.1, which will start a standalone cluster of Spark. The scripts `run_spark_cpu.sh` and `run_spark_gpu.sh` start Spark, then runs several PySpark jobs with `spark_data_utils.py`, for example:\ngenerates the dictionary\n- transforms the train dataset\n- transforms the test dataset\n- transforms the validation dataset\n\n Change the variables in the `run-spark.sh` script according to your environment.\n Configure the paths.\n```\nexport SPARK_LOCAL_DIRS=/data/spark-tmp\nexport INPUT_PATH=/data/criteo\nexport OUTPUT_PATH=/data/output\n```\nNote that the Spark job requires about 3TB disk space used for data shuffle.\n\nWhere:\n`SPARK_LOCAL_DIRS` is the path where Spark uses to write shuffle data.\n`INPUT_PATH` is the path of the Criteo Terabyte Dataset, including uncompressed files like day_0, day_1…\n`OUTPUT_PATH` is where the script writes the output data. It will generate the following subdirectories of `models`, `train`, `test`, and `validation`.\n- The `model` is the dictionary folder.\n- The `train` is the train dataset transformed from day_0 to day_22.\n- The `test` is the test dataset transformed from the prior half of day_23.\n- The `validation` is the dataset transformed from the latter half of day_23.\n\nConfigure the resources which Spark will use.\n```\nexport TOTAL_CORES=80\nexport TOTAL_MEMORY=800\n```\n\nWhere:\n`TOTAL_CORES` is the total CPU cores you want Spark to use.\n\n`TOTAL_MEMORY` is the total memory Spark will use.\n\nConfigure frequency limit.\n```\nUSE_FREQUENCY_LIMIT=15\n```\nThe frequency limit is used to filter out the categorical values which appear less than n times in the whole dataset, and make them be 0. Change this variable to 1 to enable it. The default frequency limit is 15 in the script. You also can change the number as you want by changing the line of `OPTS=\"--frequency_limit 8\"`.\n\n### Training process\n\nThe main training script resides in `main.py`. The speed of training is measured by throughput i.e., the number\nof samples processed per second. We use mixed precision training with static loss scaling for the bottom and top MLPs while embedding tables are stored in FP32 format.","hasAdvanced","performance","The performance measurements in this document were conducted at the time of publication and may not reflect the performance achieved from NVIDIA's latest software release. For the most up-to-date performance measurements, go to [NVIDIA Data Center Deep Learning Product Performance](https://developer.nvidia.com/deep-learning-performance-training-inference).\n\n### Benchmarking\n\nThe following section shows how to run benchmarks measuring the model performance in training and inference modes.\n\n#### Training performance benchmark\n\nTo benchmark the training performance on a specific batch size, follow the instructions\nin the [Quick Start Guide](#quick-start-guide). You can also add the `--max_steps 1000`\nif you want to get a reliable throughput measurement without running the entire training.\n\nYou can also use synthetic data by running with the `--dataset_type synthetic` option if you haven't downloaded the dataset yet.\n\n#### Inference performance benchmark\n\nTo benchmark the inference performance on a specific batch size, run:\n\n```\nhorovodrun -np 1 -H localhost:1 --mpi-args=--oversubscribe numactl --interleave=all -- python -u main.py --dataset_path /data/dlrm/ --amp --restore_checkpoint_path --mode inference\n```\n\n### Results\n\nThe following sections provide details on how we achieved our performance and accuracy in training and inference.\n\nWe used three model size variants to show memory scalability in multi-GPU setup:\n\n| Name | Dataset | Number of parameters | Model size |\n|:-----|:--------|:---------------------|:------|\n| small | Criteo 1TB, FL=15| 4.2B | 15.6 GiB |\n| large | Criteo 1TB, FL=3 | 22.8B | 84.9 GiB | \n| extra large | Criteo 1TB, FL=0 | 113B | 421 GiB |\n\n#### Training accuracy results\n\n##### Training accuracy: NVIDIA DGX A100 (8x A100 80GB)\n\nOur results were obtained by running training scripts as described in the Quick Start Guide in the DLRM Docker container.\n\n| GPUs | Model size | Batch size / GPU | Accuracy (AUC) - TF32 | Accuracy (AUC) - mixed precision | Time to train - TF32 [minutes] | Time to train - mixed precision [minutes] | Time to train speedup (TF32 to mixed precision) |\n|:-------|:-------------|:-------------------|:------------------------|:-----------------------------------|:---------------------------------|:--------------------------------------------|:--------------------------------------------------|\n| 1 | small | 64k | 0.8025 | 0.8025 | 26.75 | 16.27 | 1.64 |\n| 8 | large | 8k | 0.8027 | 0.8026 | 8.77 | 6.57 | 1.33 |\n| 8 | extra large | 8k | 0.8026 | 0.8026 | 10.47 | 9.08 | 1.15 |\n\n##### Training accuracy: NVIDIA DGX-1 (8x V100 32GB)\n\nOur results were obtained by running training scripts as described in the Quick Start Guide in the DLRM Docker container.\n\n| GPUs | Model size | Batch size / GPU | Accuracy (AUC) - FP32 | Accuracy (AUC) - mixed precision | Time to train - FP32 [minutes] | Time to train - mixed precision [minutes] | Time to train speedup (FP32 to mixed precision) |\n|:-------|:-------------|:-------------------|:------------------------|:-----------------------------------|:---------------------------------|:--------------------------------------------|:--------------------------------------------------|\n| 1 | small | 64k | 0.8027 | 0.8025 | 109.63 | 34.83 | 3.15 |\n| 8 | large | 8k | 0.8028 | 0.8026 | 26.01 | 13.73 | 1.89 |\n##### Training accuracy: NVIDIA DGX-2 (16x V100 32GB)\n\nOur results were obtained by running training scripts as described in the Quick Start Guide in the DLRM Docker container.\n\n| GPUs | Model size | Batch size / GPU | Accuracy (AUC) - FP32 | Accuracy (AUC) - mixed precision | Time to train - FP32 [minutes] | Time to train - mixed precision [minutes] | Time to train speedup (FP32 to mixed precision) |\n|:-------|:-------------|:-------------------|:------------------------|:-----------------------------------|:---------------------------------|:--------------------------------------------|:--------------------------------------------------|\n| 1 | small | 64k | 0.8026 | 0.8026 | 105.13 | 33.37 | 3.15 |\n| 8 | large | 8k | 0.8027 | 0.8027 | 21.21 | 11.43 | 1.86 |\n| 16 | large | 4k | 0.8025 | 0.8026 | 15.52 | 10.88 | 1.43 |\n\n##### Training stability test\n\nThe histograms below show the distribution of ROC AUC results achieved at the end of the training for each precision/hardware platform tested. There are no statistically significant differences between precision, number of GPUs or hardware platform. Using the larger dataset has a modest, positive impact on final AUC score. \n\n

\n \n
\nFigure 4. Results of stability tests for DLRM.\n

\n\n#### Training performance results\n\nWe used throughput in items processed per second as the performance metric.\n\n##### Training performance: NVIDIA DGX A100 (8x A100 80GB)\n\nOur results were obtained by following the commands from the Quick Start Guide\nin the DLRM Docker container on NVIDIA DGX A100 (8x A100 80GB) GPUs. Performance numbers (in items per second) were averaged over 1000 training steps.\n\n| GPUs | Model size | Batch size / GPU | Throughput - TF32 | Throughput - mixed precision | Throughput speedup (TF32 to mixed precision) |\n|:-------|:-------------|:-------------------|:--------------------|:-------------------------------|:-----------------------------------------------|\n| 1 | small | 64k | 2.68M | 4.47M | 1.67 |\n| 8 | large | 8k | 9.39M | 13.31M | 1.42 |\n| 8 | extra large | 8k | 9.93M | 12.1M | 1.22\n\nTo achieve these results, follow the steps in the [Quick Start Guide](#quick-start-guide).\n\n##### Training performance: comparison with CPU for the \"extra large\" model \n\nFor the \"extra large\" model (113B parameters) we also obtained CPU results for comparison using the same source code\n(using the `--cpu` command line flag for the CPU-only experiments). \n\nWe compare three hardware setups:\n- CPU only,\n- a single GPU that uses CPU memory for the largest embedding tables,\n- Hybrid-Parallel using the full DGX A100-80GB\n\n| Hardware | Throughput [samples / second]| Speedup over CPU|\n|:---|:---|:---|\n2xAMD EPYC 7742 | 17.7k | 1x |\n1xA100-80GB + 2xAMD EPYC 7742 (large embeddings on CPU) | 768k |43x |\nDGX A100 (8xA100-80GB) (hybrid parallel) | 12.1M | 683x |\n\n##### Training performance: NVIDIA DGX-1 (8x V100 32GB)\n\n| GPUs | Model size | Batch size / GPU | Throughput - FP32 | Throughput - mixed precision | Throughput speedup (FP32 to mixed precision) |\n|:-------|:-------------|:-------------------|:--------------------|:-------------------------------|:-----------------------------------------------|\n| 1 | small | 64k | 0.648M | 2.06M | 3.18 |\n| 8 | large | 8k | 2.9M | 5.89M | 2.03 |\n\nTo achieve the same results, follow the steps in the [Quick Start Guide](#quick-start-guide).\n\n##### Training performance: NVIDIA DGX-2 (16x V100 32GB)\n\n| GPUs | Model size | Batch size / GPU | Throughput - FP32 | Throughput - mixed precision | Throughput speedup (FP32 to mixed precision) |\n|:-------|:-------------|:-------------------|:--------------------|:-------------------------------|:-----------------------------------------------|\n| 1 | small | 64k | 0.675M | 2.16M | 3.2 |\n| 8 | large | 8k | 3.75M | 7.72M | 2.06 |\n| 16 | large | 4k | 5.74M | 9.39M | 1.64 |\n\nTo achieve the same results, follow the steps in the [Quick Start Guide](#quick-start-guide).\n\n#### Inference performance results\n\n##### Inference performance: NVIDIA DGX A100 (8x A100 80GB)\n\n| GPUs | Model size | Batch size / GPU | Throughput - TF32 | Throughput - mixed precision | Average latency - TF32 [ms] | Average latency - mixed precision [ms] | Throughput speedup (mixed precision to TF32) |\n|-------:|:-------------|-------------------:|:--------------------|:-------------------------------|------------------------------:|-----------------------------------------:|-----------------------------------------------:|\n| 1 | small | 2048 | 1.43M | 1.54M | 1.48 | 1.33 | 1.08 |\n\n##### Inference performance: NVIDIA DGX1V-32GB (8x V100 32GB)\n\n| GPUs | Model size | Batch size / GPU | Throughput - FP32 | Throughput - mixed precision | Average latency - FP32 [ms] | Average latency - mixed precision [ms] | Throughput speedup (mixed precision to FP32) |\n|-------:|:-------------|-------------------:|:--------------------|:-------------------------------|------------------------------:|-----------------------------------------:|-----------------------------------------------:|\n| 1 | small | 2048 | 0.765M | 1.05M | 2.90 | 1.95 | 1.37 |\n\n##### Inference performance: NVIDIA DGX2 (16x V100 16GB)\n\n| GPUs | Model size | Batch size / GPU | Throughput - FP32 | Throughput - mixed precision | Average latency - FP32 [ms] | Average latency - mixed precision [ms] | Throughput speedup (mixed precision to FP32) |\n|-------:|:-------------|-------------------:|:--------------------|:-------------------------------|------------------------------:|-----------------------------------------:|-----------------------------------------------:|\n| 1 | small | 2048 | 1.03M | 1.37M | 2.10 | 1.63 | 1.53 |","hasPerformance",[],{"_86":233,"_113":234,"_114":-5,"_122":34,"_121":34,"_123":232,"_216":91},{"_86":28,"_113":28,"_114":-5,"_122":34,"_121":34,"_123":232,"_216":91},[],"-","Latest (22.06.0)","nvidia/dle/dlrm_tf2","key","nvidia/dle/resources/dlrm_tf2","resourceId","dlrm_tf2","shortDescription","The Deep Learning Recommendation Model (DLRM) is a recommendation model designed to make use of both categorical and numerical inputs.","This resource is using open-source code maintained in github (see the quick-start-guide section) and available for download from [NGC](https://catalog.ngc.nvidia.com/orgs/nvidia/teams/dle/resources/dlrm-tf2)\n\n\nThe Deep Learning Recommendation Model (DLRM) is a recommendation model designed to make use of both categorical and numerical inputs.\nIt was first described in [Deep Learning Recommendation Model for Personalization and Recommendation Systems](https://arxiv.org/abs/1906.00091).\nThis repository provides a reimplementation of the code-base provided originally [here](https://github.com/facebookresearch/dlrm).\nThe scripts enable you to train DLRM on the [Criteo Terabyte Dataset](https://labs.criteo.com/2013/12/download-terabyte-click-logs/).\n\nUsing the scripts provided here, you can efficiently train models that are too large to fit into a single GPU.\nThis is because we use a hybrid-parallel approach, which combines model parallelism with data parallelism for\ndifferent parts of the neural network.\nThis is explained in details in the [next section](#hybrid-parallel-multi-gpu-with-all-2-all-communication).\n\nThis model uses a slightly different preprocessing procedure than the one found in the original implementation.\nMost importantly, we use a technique called frequency thresholding to demonstrate models of different size.\nThe smallest model can be trained on a single V100-32GB GPU, while the largest one needs 8xA100-80GB GPUs. \nThe table below summarizes the model sizes and frequency thresholds used in this repository:\n\n| Name | Frequency threshold | Number of parameters | Model size|\n|:--------------|:---|:-------------|:-------------------|\n| Small | 15| 4.2B | 15.6 GiB |\n| Large | 3 | 22.8B | 84.9 GiB |\n| Extra large | 0 | 113B | 421 GiB |\n\nYou can find a detailed description of the preprocessing steps in the [Dataset guidelines](#dataset-guidelines) section.\n\nUsing DLRM, you can train a high-quality general model for recommendations.\n\nThis model is trained with mixed precision using Tensor Cores on Volta, Turing and NVIDIA Ampere GPU architectures.\nTherefore, researchers can get results 2x faster than training without Tensor Cores while experiencing the\nbenefits of mixed precision training. This model is tested against each NGC monthly container\nrelease to ensure consistent accuracy and performance over time.\n\n### Model architecture\n\nDLRM accepts two types of features: categorical and numerical. For each categorical feature, an embedding table is used to provide dense representation to each unique value. The dense features enter the model and are transformed by a simple neural network referred to as \"bottom MLP\".\n\nThis part of the network consists of a series\nof linear layers with ReLU activations. The output of the bottom MLP and the embedding vectors are then fed into the \"dot interaction\" operation. The output of \"dot interaction\" is then concatenated with the features resulting from bottom MLP and fed into the \"top MLP\" which is a series of dense layers with activations.\nThe model outputs a single number which can be interpreted as a likelihood of a certain user clicking an ad.\n\n

\n \n
\nFigure 1. The architecture of DLRM.\n

\n\n### Default configuration\n\nThe following features were implemented in this model:\n- general\n\t- static loss scaling for Tensor Cores (mixed precision) training\n\t- hybrid-parallel multi-GPU training\n- preprocessing\n - dataset preprocessing using Spark 3 on GPUs\n\n### Feature support matrix\n\nThe following features are supported by this model:\n\n| Feature | DLRM\n|----------------------|--------------------------\n|Automatic mixed precision (AMP) | Yes\n|XLA | Yes\n|Hybrid-parallel training with Merlin Distributed Embeddings | Yes\n|Preprocessing on GPU with Spark 3| Yes\n|Multi-node training | Yes\n\n#### Features\n\n**Automatic Mixed Precision (AMP)**\nEnables mixed precision training without any changes to the code-base by performing automatic graph rewrites and loss scaling controlled by an environmental variable.\n\n**XLA**\n\nThe training script supports a `--xla` flag. It can be used to enable XLA JIT compilation. Currently, we use [XLA Lite](https://docs.nvidia.com/deeplearning/frameworks/tensorflow-user-guide/index.html#xla-lite). It delivers a steady 10-30% performance boost depending on your hardware platform, precision, and the number of GPUs. It is turned off by default.\n\n**Horovod**\nHorovod is a distributed training framework for TensorFlow, Keras, PyTorch, and MXNet. The goal of Horovod is to make distributed deep learning fast and easy to use. For more information about how to get started with Horovod, see the Horovod [official repository](https://github.com/horovod/horovod).\n\n**Hybrid-parallel training with Merlin Distributed Embeddings**\nOur model uses Merlin Distributed Embeddings to implement efficient multi-GPU training.\nFor details, see example sources in this repository or see the TensorFlow tutorial.\nFor the detailed description of our multi-GPU approach, visit this [section](#hybrid-parallel-training-with-merlin-distributed-embeddings).\n\n**Multi-node training**\nThis repository supports multinode training. For more information refer to the [multinode section](#multi-node-training)\n\n### Mixed precision training\n\nMixed precision is the combined use of different numerical precisions in a computational method. [Mixed precision](https://arxiv.org/abs/1710.03740) training offers significant computational speedup by performing operations in half-precision format while storing minimal information in single-precision to retain as much information as possible in critical parts of the network. Since the introduction of [Tensor Cores](https://developer.nvidia.com/tensor-cores) in Volta, and following with both the Turing and Ampere architectures, significant training speedups are experienced by switching to mixed precision -- up to 3.4x overall speedup on the most arithmetically intense model architectures. Using mixed precision training requires two steps:\n1. Porting the model to use the FP16 data type where appropriate.\n2. Adding loss scaling to preserve small gradient values.\n\nThe ability to train deep learning networks with lower precision was introduced in the Pascal architecture and first supported in [CUDA 8](https://devblogs.nvidia.com/parallelforall/tag/fp16/) in the NVIDIA Deep Learning SDK.\n\nFor information about:\n- How to train using mixed precision, see the [Mixed Precision Training](https://arxiv.org/abs/1710.03740) paper and [Training With Mixed Precision](https://docs.nvidia.com/deeplearning/performance/mixed-precision-training/index.html) documentation.\n- Techniques used for mixed precision training, see the [Mixed-Precision Training of Deep Neural Networks](https://devblogs.nvidia.com/mixed-precision-training-deep-neural-networks/) blog.\n\n#### Enabling mixed precision\n\nMixed precision training is turned off by default. To turn it on, issue the `--amp` flag to the `main.py` script.\n\n#### Enabling TF32\n\nTensorFloat-32 (TF32) is the new math mode in [NVIDIA A100](https://www.nvidia.com/en-us/data-center/a100/) GPUs for handling the matrix math also called tensor operations. TF32 running on Tensor Cores in A100 GPUs can provide up to 10x speedups compared to single-precision floating-point math (FP32) on Volta GPUs.\n\nTF32 Tensor Cores can speed up networks using FP32, typically with no loss of accuracy. It is more robust than FP16 for models which require high dynamic range for weights or activations.\n\nFor more information, refer to the [TensorFloat-32 in the A100 GPU Accelerates AI Training, HPC up to 20x](https://blogs.nvidia.com/blog/2020/05/14/tensorfloat-32-precision-format/) blog post.\n\nTF32 is supported in the NVIDIA Ampere GPU architecture and is enabled by default.\n\n### Hybrid-parallel training with Merlin Distributed Embeddings\n\nMany recommendation models contain very large embedding tables. As a result, the model is often too large to fit onto a single device.\nThis could be easily solved by training in a model-parallel way, using either the CPU or other GPUs as \"memory donors\".\nHowever, this approach is suboptimal as the \"memory donor\" devices' compute is not utilized.\nIn this repository, we use the model-parallel approach for the Embedding Tables while employing a usual data parallel approach\nfor the more compute-intensive MLPs and Dot Interaction layer. This way, we can train models much larger than what would normally fit into\na single GPU while at the same time making the training faster by using multiple GPUs. We call this approach hybrid-parallel training.\n\nTo implement this approach we use the [Merlin Distributed Embeddings](https://github.com/NVIDIA-Merlin/distributed-embeddings) library. \nIt provides a scalable model parallel wrapper called `distributed_embeddings.dist_model_parallel`. This wrapper automatically distributes embedding tables to multiple GPUs.\nThis way embeddings can be scaled beyond single GPU's memory capacity without\ncomplex code to handle cross-worker communication.\n\nUnder the hood, Merlin Distributed Embeddings uses a\nspecific multi-GPU communication pattern called\n[all-2-all](https://en.wikipedia.org/wiki/All-to-all_\$parallel_pattern\$) to transition from model-parallel to data-parallel\nparadigm. In the [original DLRM whitepaper](https://arxiv.org/abs/1906.00091) this has been referred to as \"butterfly shuffle\".\n\nAn example model using Hybrid Parallelism is shown in Figure 2. The compute intensive dense layers are run in data-parallel\nmode. The smaller embedding tables are run model-parallel, such that each smaller table is placed entirely on a single device.\nThis is not suitable for larger tables that need more memory than can be provided by a single device. Therefore,\nthose large tables are split into multiple parts and each part is run on a different GPU.\n\n

\n \n
\nFigure 2. Hybrid parallelism with Merlin Distributed Embeddings.\n

\n\nIn this repository we train models of three sizes: \"small\" (15.6 GiB), \"large\" (84.9 GiB) and \"extra large\" (421 GiB).\nThe \"small\" model can be trained on a single V100-32GB GPU. The \"large\" model needs at least 8xV100-32GB GPUs,\nbut each of the tables it uses can fit on a singleGPU. \n\nThe \"extra large\" model, on the other hand, contains tables that do not fit into a singledevice, and will be automatically\nsplit and stored across multiple GPUs by Merlin Distributed Embeddings.\n\n#### Training very large embedding tables\n\nWe tested this approach by training a DLRM model on the Criteo Terabyte dataset with the frequency limiting option turned off (set to zero).\nThe weights of the resulting model take 421 GiB. The largest table weighs 140 GiB.\nHere are the commands you can use to reproduce this:\n\n```\n# build and run the preprocessing container as in the Quick Start Guide\n# then when preprocessing set the frequency limit to 0:\n./prepare_dataset.sh DGX2 0\n\n# build and run the training container same as in the Quick Start Guide\n# then append options necessary for training very large embedding tables:\nhorovodrun -np 8 -H localhost:8 --mpi-args=--oversubscribe numactl --interleave=all -- python -u main.py --dataset_path /data/dlrm/ --amp --xla\n```\n\nWhen using this method on a DGX A100 with 8 A100-80GB GPUs and a large-enough dataset, it is possible to train a single embedding table of up to 600 GB. You can also use multi-node training (described below) to train even larger recommender systems.\n\nThis mode was used to train the 421GiB \"extra large\" model in [the DGX A100-80G performance section](#training-performance-nvidia-dgx-a100-8x-a100-80gb).\n\n#### Multi-node training\n\nMulti-node training is supported. Depending on the exact interconnect hardware and model configuration,\nyou might experience only a modest speedup with multi-node.\nMulti-node training can also be used to train larger models.\nFor example, to train a 1.68 TB variant of DLRM on multi-node, you can run:\n\n```\ncmd='numactl --interleave=all -- python -u main.py --dataset_path /data/dlrm/full_criteo_data --amp --xla\\\n--embedding_dim 512 --bottom_mlp_dims 512,256,512' \\\nsrun_flags='--mpi=pmix' \\\ncont=nvidia_dlrm_tf \\\nmounts=/data/dlrm:/data/dlrm \\\nsbatch -n 32 -N 4 -t 00:20:00 slurm_multinode.sh\n```\n\n### Preprocessing on GPU with Spark 3\n\nRefer to the [\"Preprocessing with Spark\" section](#preprocess-with-spark) for a detailed description of the Spark 3 GPU functionality.\n\n### BYO dataset functionality overview\n\nThis section describes how you can train the DeepLearningExamples RecSys models on your own datasets without changing\nthe model or data loader and with similar performance to the one published in each repository.\nThis can be achieved thanks to Dataset Feature Specification, which describes how the dataset, data loader and model\ninteract with each other during training, inference and evaluation.\nDataset Feature Specification has a consistent format across all recommendation models in NVIDIA's DeepLearningExamples\nrepository, regardless of dataset file type and the data loader,\ngiving you the flexibility to train RecSys models on your own datasets.\n\n- [Glossary](#glossary)\n- [Dataset Feature Specification](#dataset-feature-specification)\n- [Data Flow in Recommendation Models in DeepLearning examples](#data-flow-in-nvidia-deep-learning-examples-recommendation-models)\n- [Example of Dataset Feature Specification](#example-of-dataset-feature-specification)\n- [BYO dataset functionality](#byo-dataset-functionality)\n\n#### Glossary\n\nThe Dataset Feature Specification consists of three mandatory and one optional section:\n\nfeature_spec provides a base of features that may be referenced in other sections, along with their metadata.\n\tFormat: dictionary (feature name) => (metadata name => metadata value)
\n\nsource_spec provides information necessary to extract features from the files that store them. \n\tFormat: dictionary (mapping name) => (list of chunks)
\n\n* Mappings are used to represent different versions of the dataset (think: train/validation/test, k-fold splits). A mapping is a list of chunks.
\n* Chunks are subsets of features that are grouped together for saving. For example, some formats may constrain data saved in one file to a single data type. In that case, each data type would correspond to at least one chunk. Another example where this might be used is to reduce file size and enable more parallel loading. Chunk description is a dictionary of three keys:
\n * type provides information about the format in which the data is stored. Not all formats are supported by all models.
\n * features is a list of features that are saved in a given chunk. Order of this list may matter: for some formats, it is crucial for assigning read data to the proper feature.
\n * files is a list of paths to files where the data is saved. For Feature Specification in yaml format, these paths are assumed to be relative to the yaml file's directory (basename). Order of this list matters: It is assumed that rows 1 to i appear in the first file, rows i+1 to j in the next one, etc.
\n\nchannel_spec determines how features are used. It is a mapping (channel name) => (list of feature names). \n\nChannels are model specific magic constants. In general, data within a channel is processed using the same logic. Example channels: model output (labels), categorical ids, numerical inputs, user data, and item data.\n\nmetadata is a catch-all, wildcard section: If there is some information about the saved dataset that does not fit into the other sections, you can store it here.\n\n#### Dataset feature specification\n\nData flow can be described abstractly:\nInput data consists of a list of rows. Each row has the same number of columns; each column represents a feature.\nThe columns are retrieved from the input files, loaded, aggregated into channels and supplied to the model/training script. \n\nFeatureSpec contains metadata to configure this process and can be divided into three parts:\n\n* Specification of how data is organized on disk (source_spec). It describes which feature (from feature_spec) is stored in which file and how files are organized on disk.\n\n* Specification of features (feature_spec). Describes a dictionary of features, where key is feature name and values are features' characteristics such as dtype and other metadata (for example, cardinalities for categorical features)\n\n* Specification of model's inputs and outputs (channel_spec). Describes a dictionary of model's inputs where keys specify model channel's names and values specify lists of features to be loaded into that channel. Model's channels are groups of data streams to which common model logic is applied, for example categorical/continuous data, user/item ids. Required/available channels depend on the model\n\nThe FeatureSpec is a common form of description regardless of underlying dataset format, dataset data loader form and model. \n\n#### Data flow in NVIDIA Deep Learning Examples recommendation models\n\nThe typical data flow is as follows:\n* S.0. Original dataset is downloaded to a specific folder.\n* S.1. Original dataset is preprocessed into Intermediary Format. For each model, the preprocessing is done differently, using different tools. The Intermediary Format also varies (for example, for DLRM PyTorch, the Intermediary Format is a custom binary one.)\n* S.2. The Preprocessing Step outputs Intermediary Format with dataset split into training and validation/testing parts along with the Dataset Feature Specification yaml file. Metadata in the preprocessing step is automatically calculated.\n* S.3. Intermediary Format data together with Dataset Feature Specification are fed into training/evaluation scripts. Data loader reads Intermediary Format and feeds the data into the model according to the description in the Dataset Feature Specification.\n* S.4. The model is trained and evaluated\n\n

\n \n
\n\nFig.1. Data flow in Recommender models in NVIDIA Deep Learning Examples repository. Channels of the model are drawn in green.\n

\n\n#### Example of dataset feature specification\n\nAs an example, let's consider a Dataset Feature Specification for a small CSV dataset for some abstract model.\n\n```yaml\nfeature_spec:\n user_gender:\n dtype: torch.int8\n cardinality: 3 #M,F,Other\n user_age: #treated as numeric value\n dtype: torch.int8\n user_id:\n dtype: torch.int32\n cardinality: 2655\n item_id:\n dtype: torch.int32\n cardinality: 856\n label:\n dtype: torch.float32\n\nsource_spec:\n train:\n - type: csv\n features:\n - user_gender\n - user_age\n files:\n - train_data_0_0.csv\n - train_data_0_1.csv\n - type: csv\n features:\n - user_id\n - item_id\n - label\n files:\n - train_data_1.csv\n test:\n - type: csv\n features:\n - user_id\n - item_id\n - label\n - user_gender\n - user_age\n \n files:\n - test_data.csv\n\nchannel_spec:\n numeric_inputs: \n - user_age\n categorical_user_inputs: \n - user_gender\n - user_id\n categorical_item_inputs: \n - item_id\n label_ch: \n - label\n```\n\nThe data contains five features: (user_gender, user_age, user_id, item_id, label). Their data types and necessary metadata are described in the feature specification section.\n\nIn the source mapping section, two mappings are provided: one describes the layout of the training data, the other of the testing data. The layout for training data has been chosen arbitrarily to showcase the flexibility.\nThe train mapping consists of two chunks. The first one contains user_gender and user_age, saved as a CSV, and is further broken down into two files. For specifics of the layout, refer to the following example and consult the glossary. The second chunk contains the remaining columns and is saved in a single file. Notice that the order of columns is different in the second chunk - this is alright, as long as the order matches the order in that file (that is, columns in the .csv are also switched)\n\nLet's break down the train source mapping. The table contains example data color-paired to the files containing it.\n\n

\n\n

\n\nThe channel spec describes how the data will be consumed. Four streams will be produced and available to the script/model.\nThe feature specification does not specify what happens further: names of these streams are only lookup constants defined by the model/script.\nBased on this example, we can speculate that the model has three input channels: numeric_inputs, categorical_user_inputs,\ncategorical_item_inputs, and one output channel: label.\nFeature names are internal to the FeatureSpec and can be freely modified.\n\n#### BYO dataset functionality\n\nIn order to train any Recommendation model in NVIDIA Deep Learning Examples one can follow one of three possible ways:\n* One delivers already preprocessed dataset in the Intermediary Format supported by data loader used by the training script\n(different models use different data loaders) together with FeatureSpec yaml file describing at least specification of dataset, features and model channels\n\n* One uses a transcoding script\n\n* One delivers dataset in non-preprocessed form and uses preprocessing scripts that are a part of the model repository.\nIn order to use already existing preprocessing scripts, the format of the dataset needs to match the one of the original datasets.\nThis way, the FeatureSpec file will be generated automatically, but the user will have the same preprocessing as in the original model repository.","displayName","DLRM for TensorFlow2",["D",1680631656496],"publisher","NVIDIA Deep Learning Examples","latestVersion","collectionCounts","imageUrl","/api/image/nvidia/dle/resources/dlrm_tf2","rawImageUrl","https://assets.nvidiagrid.net/ngc/logos/Deep-Learning-Examples.png","canGuestDownload","latestVersionSizeInBytes","licenseTerms","isMultinodeEnabled","isMultiArchitecture","bias","explainability","privacy","safetyAndSecurity","featureLabels",[],"displayLabels",[274],"policyLabels",[],"systemLabels",[],"products",[],"hasNimMetadata",{"_243":275,"_276":277,"_278":279,"_280":88},"Recommendation","value","uscs_recommedation","filter",{"_281":282,"_276":277},"category","field","application",{"_152":1261,"_154":1262},"ngcApps",[1214],"csaApiBaseUrl","https://www.nvidia.com/ai-agents/csa-agent/dcsm","datadogRum",{"_1195":1196,"_1197":1198,"_1199":1200,"_1201":1202,"_1203":294,"_110":320,"_1204":1205,"_1206":59,"_1207":91,"_1208":91,"_1209":1210,"_1211":91,"_1212":1213},"amplitude",{"_1192":1193,"_833":-7,"_1194":34},"disableJs","environment","prod","featureFlagsBlob","JTdCJTIydGVzdEZsYWclMjIlM0FmYWxzZSUyQyUyMnNob3dTZWFyY2hMYWJlbFNldEZpbHRlcnMlMjIlM0FmYWxzZSUyQyUyMmVuYWJsZUNzYVdpZGdldCUyMiUzQWZhbHNlJTJDJTIyZW5hYmxlUHJvZHVjdEdyb3VwaW5nQnlTdWJzY3JpcHRpb24lMjIlM0FmYWxzZSUyQyUyMmNvbnRhaW5lclNjYW5uaW5nUmVkZXNpZ25NMiUyMiUzQWZhbHNlJTdE","featureFlagsGuarded","isDevLikeEnv","isPlaywright","isMobileDevice","categorizedOrgs",{"_1186":1187,"_1188":1189,"_1190":1191},"currentOrgContext","currentContextRoles",[],"enterpriseOrgs",[],"manageableExternalUserGroups",[],"ngcApiOrigin","ngcHelmOrigin","origin","https://catalog.ngc.nvidia.com","usesMocks","theme","year",2026,"now",1787557791797,"1.12.0","defaultSearchSuggestions",{"_1073":401,"_1074":1075,"_1076":1077,"_1078":1079},"announcements",[],"consentState",{"_358":1063,"_1064":91,"_1065":34,"_1066":34,"_1067":1068},"consentConfig",{"_340":341,"_342":343,"_344":345,"_346":347,"_348":349,"_350":351,"_352":353,"_354":355,"_356":357,"_358":359},"showContactAdmin",1787557791802,[],[],[334,335,336,337,338,339],{"_48":63,"_50":74,"_52":147,"_54":-5,"_56":-5,"_58":59},{"_48":63,"_50":70,"_52":147,"_54":-5,"_56":-5,"_58":59},{"_48":63,"_50":68,"_52":147,"_54":-5,"_56":-5,"_58":59},{"_48":63,"_50":64,"_52":147,"_54":-5,"_56":-5,"_58":59},{"_48":60,"_50":61,"_52":147,"_54":-5,"_56":-5,"_58":59},{"_48":49,"_50":51,"_52":147,"_54":-5,"_56":-5,"_58":59},"mainInfoText","We and our third-party partners (including social media, advertising, and analytics partners) use cookies and other tracking technologies to collect, store, monitor, and process certain information about you when you visit our website. The information collected might relate to you, your preferences, or your device. We use that information to make the site work, analyze performance and traffic on our website, provide a more personalized web experience, and assist in our marketing efforts.

Under certain privacy laws, you have the right to direct us not to \"sell\" or \"share\" your personal information for targeted advertising. To opt-out of the \"sale\" and \"sharing\" of personal information through cookies, you must opt-out of optional cookies using the toggles below. To opt out of the \"sale\" and \"sharing\" of data collected by other means (e.g., online forms) you must also update your data sharing preferences through the NVIDIA Preference Center.

Click on the different category headings below to find out more and change the settings according to your preference. You cannot opt out of Required Cookies as they are deployed to ensure the proper functioning of our website (such as prompting the cookie banner and remembering your settings, etc.). By clicking \"Save and Accept\" or \"Decline All\" at the bottom, you consent to the use of cookies and other tools as described in our Cookie Policy in accordance with your settings and accept our Terms of Service (which contains important waivers). For more information about our privacy practices, please see our Privacy Policy.","mainText","Cookie Settings","alertNoticeText",{"_1055":1056,"_1057":1058,"_1059":1060,"_1061":1062},"bannerAcceptLabel","Accept All","bannerRejectLabel","Reject Optional","bannerSettingsLabel","Manage Settings","preferenceSaveLabel","Save Choices","preferenceRejectLabel","Decline All","alwaysActiveText","Always Active","groups",[360,361,362,363,364],{"_365":745,"_113":746,"_210":747,"_369":91,"_370":16,"_372":748,"_374":749},{"_365":648,"_113":649,"_210":650,"_369":34,"_370":57,"_372":651,"_374":652},{"_365":556,"_113":557,"_210":558,"_369":34,"_370":62,"_372":559,"_374":560},{"_365":376,"_113":377,"_210":378,"_369":34,"_370":379,"_372":380,"_374":381},{"_365":366,"_113":367,"_210":368,"_369":34,"_370":371,"_372":373,"_374":375},"optanonGroupId","C0005","Social Media Cookies","These cookies are set by a range of social media services that we have added to the site to enable you to share our content with your friends and networks. They are capable of tracking your browser across other sites and building up a profile of your interests. This may impact the content and messages you see on other websites you visit. If you do not allow these cookies you may not be able to use or see these sharing tools.","alwaysActive","order",11,"cookies",[],"hosts",[],"C0004","Advertising Cookies","These cookies record your visit to our websites, the pages you have visited and the links you have followed to influence the advertisements that you see on other websites. These cookies and the information they collect may be managed by other companies, including our advertising partners, and may be used to build a profile of your interests and show you relevant advertising on other sites. We and our advertising partners will use this information to make our websites and the advertising displayed on it, more relevant to your interests.",5,[498,499,500,501,502,503,504,505,506,507,508,509,510,511,512,513,514,515,516,517,518,519],[382,383,384,385,386,387,388,389,390,391,392,393,394,395,396,397,398],{"_399":493,"_243":493,"_210":401,"_372":494},{"_399":489,"_243":489,"_210":401,"_372":490},{"_399":484,"_243":484,"_210":401,"_372":485},{"_399":480,"_243":480,"_210":401,"_372":481},{"_399":475,"_243":475,"_210":401,"_372":476},{"_399":470,"_243":470,"_210":401,"_372":471},{"_399":462,"_243":462,"_210":401,"_372":463},{"_399":454,"_243":454,"_210":401,"_372":455},{"_399":449,"_243":449,"_210":401,"_372":450},{"_399":444,"_243":444,"_210":401,"_372":445},{"_399":439,"_243":439,"_210":401,"_372":440},{"_399":434,"_243":434,"_210":401,"_372":435},{"_399":429,"_243":429,"_210":401,"_372":430},{"_399":422,"_243":422,"_210":401,"_372":423},{"_399":417,"_243":417,"_210":401,"_372":418},{"_399":410,"_243":410,"_210":401,"_372":411},{"_399":400,"_243":400,"_210":401,"_372":402},"hostName","amazon-adsystem.com","",[403,404],{"_113":409,"_406":400,"_407":34,"_210":408},{"_113":405,"_406":400,"_407":34,"_210":408},"ad-privacy","host","isSession","This domain is owned by online retailer Amazon and is used as part of its affiliate marketing programme.","ad-id","doubleclick.net",[412,413],{"_113":416,"_406":410,"_407":34,"_210":415},{"_113":414,"_406":410,"_407":34,"_210":415},"IDE","This domain is owned by Doubleclick (Google). The main business activity is: Doubleclick is Googles real time bidding advertising exchange","test_cookie","bing.com",[419],{"_113":420,"_406":417,"_407":34,"_210":421},"MUID","This domain is owned by Mircosoft - it is the site for the search engine Bing.","assets.adobedtm.com",[424,425],{"_113":428,"_406":422,"_407":34,"_210":427},{"_113":426,"_406":422,"_407":34,"_210":427},"kndctr_F207D74D549850760A4C98C6_AdobeOrg_*","If you know anything about this host and its uses please add your contributions here. ","kndctr_F207D74D549850760A4C98C6_AdobeOrg_consent","tiktok.com",[431],{"_113":432,"_406":429,"_407":34,"_210":433},"_ttp","TikTok is a social media platform owned by ByteDance, known for short-form video sharing. The service is used to create and share videos, and it utilizes cookies for various purposes including advertising and analytics.","t.co",[436],{"_113":437,"_406":434,"_407":34,"_210":438},"muc_ads","This domain is owned by Twitter. It is used as a URL shortening service, which allows Twitter to collect data on how links are shared.","snowplow.com",[441],{"_113":442,"_406":439,"_407":34,"_210":443},"sp","Testing","google.com",[446],{"_113":447,"_406":444,"_407":34,"_210":448},"NID","This domain is owned by Google Inc. Although Google is primarily known as a search engine, the company provides a diverse range of products and services. Its main source of revenue however is advertising. Google tracks users extensively both through its own products and sites, and the numerous technologies embedded into many millions of websites around the world. It uses the data gathered from most of these services to profile the interests of web users and sell advertising space to organisations based on such interest profiles as well as aligning adverts to the content on the pages where its customer's adverts appear.","linkedin.com",[451],{"_113":452,"_406":449,"_407":34,"_210":453},"bcookie","This domain is owned by LinkedIn, the business networking platform. It typically acts as a third party host where website owners have placed one of its content sharing buttons in their pages, although its content and services can be embedded in other ways. Although such buttons add functionality to the website they are on, cookies are set regardless of whether or not the visitor has an active Linkedin profile, or agreed to their terms and conditions. For this reason it is classified as a primarily tracking/targeting domain.","twitter.com",[456,457],{"_113":460,"_406":454,"_407":34,"_210":461},{"_113":458,"_406":454,"_407":34,"_210":459},"personalization_id","This domain is owned by Twitter. The main business activity is: Social Networking Services. Where twitter acts as a third party host, it collects data through a range of plug-ins and integrations, that is primarily used for tracking and targeting.","guest_*","This domain is owned by X (formerly Twitter). The main business activity is: Social Networking Services. Where X (formerly Twitter) acts as a third party host, it collects data through a range of plug-ins and integrations, that is primarily used for tracking and targeting.","youtube.com",[464,465],{"_113":468,"_406":462,"_407":34,"_210":469},{"_113":466,"_406":462,"_407":34,"_210":467},"VISITOR_INFO1_LIVE","This cookie is used as a unique identifier to track viewing of videos","__Secure-*","YouTube is a Google owned platform for hosting and sharing videos. YouTube collects user data through videos embedded in websites, which is aggregated with profile data from other Google services in order to display targeted advertising to web visitors across a broad range of their own and other websites.","demdex.net",[472],{"_113":473,"_406":470,"_407":34,"_210":474},"demdex","This cookie helps Adobe Audience Manger perform basic functions such as visitor identification, ID synchronization, segmentation, modeling, reporting, etc.","im-apps.net",[477],{"_113":478,"_406":475,"_407":34,"_210":479},"receive-cookie-deprecation","This domain is owned by im-apps.net which provides information and resources about various instant messaging (IM) applications to help users make informed decisions based on their communication needs.","influ2.com",[482],{"_113":483,"_406":480,"_407":34,"_210":443},"R","6sc.co",[486],{"_113":487,"_406":484,"_407":34,"_210":488},"6suuid","This domain is associated with the company 6Sense Insights, Inc., which provides predictive analytics and account engagement services. It is primarily used for tracking and targeting in marketing automation and advertising.","hushly.com",[491],{"_113":492,"_406":489,"_407":34,"_210":443},"_hly_gvid","t.contentsquare.net",[495],{"_113":496,"_406":493,"_407":91,"_210":497},"_cs_*","This is a session cookie. This is a pattern type cookie where a random 13-digit number is appended to the prefix _cs_.",{"_113":554,"_406":521,"_407":34,"_210":555},{"_113":552,"_406":521,"_407":34,"_210":553},{"_113":550,"_406":521,"_407":34,"_210":551},{"_113":432,"_406":521,"_407":34,"_210":549},{"_113":526,"_406":521,"_407":91,"_210":528},{"_113":548,"_406":521,"_407":34,"_210":401},{"_113":547,"_406":524,"_407":34,"_210":401},{"_113":545,"_406":521,"_407":34,"_210":546},{"_113":543,"_406":521,"_407":34,"_210":544},{"_113":542,"_406":521,"_407":34,"_210":401},{"_113":540,"_406":521,"_407":34,"_210":541},{"_113":526,"_406":539,"_407":91,"_210":528},{"_113":537,"_406":524,"_407":34,"_210":538},{"_113":537,"_406":527,"_407":34,"_210":538},{"_113":496,"_406":521,"_407":34,"_210":497},{"_113":535,"_406":521,"_407":34,"_210":536},{"_113":442,"_406":533,"_407":34,"_210":534},{"_113":531,"_406":521,"_407":34,"_210":532},{"_113":529,"_406":521,"_407":34,"_210":530},{"_113":526,"_406":527,"_407":91,"_210":528},{"_113":523,"_406":524,"_407":34,"_210":525},{"_113":520,"_406":521,"_407":91,"_210":522},"AMCVS_xxxxxAdobeOrg","nvidia.com","This is a pattern type cookie name associated with Adobe Marketing Cloud. It stores a unique visitor identifier, and uses an organisation identifier to allow a company to track users across their domains and services.This is a pattern type cookie name associated with Adobe Marketing Cloud. It stores a unique visitor identifier, and uses an organisation identifier.","_rdt_pn","www.nvidia.com","This cookie name is associated with www.mind.org.uk which is used for marketing purposes.","AMP_*","build.nvidia.com","Supports Amplitude product analytics.","_gcl_au","Used by Google AdSense for experimenting with advertisement efficiency across websites using their services","_mkto_trk","This cookie is associated with an email marketing service provided by Marketo. This tracking cookie allows a website to link visitor behaviour to the recipient of an email marketing campaign, to measure campaign effectiveness.","investor.nvidia.com","This cookie is associated with Eventbrite and is used to deliver content tailored to the end user's interests and improve content creation. This cookie is also used for event-booking purposes.","_gat_gtag_UA_140825383_2","Google Analytics","_rdt_uuid","This cookie is set by Reddit and is used for tracking the views of embedded videos and performance of the advertisements.","catalog.ngc.nvidia.com","_uetsid","This cookie is used by Bing to determine what ads should be shown that may be relevant to the end user perusing the site. ","kndctr_F207D74D549850760A4C98C6_AdobeOrg_bc_session_id","ttcsid_XXXXXXXXXXXXXXXXXXXXCategory","This cookie name is associated with TikTok which is used to track your activity on the platform, including how you interact with their embedded services on other websites.This cookie is associated with TikTok. It helps track your activity on the platform, including how you interact with their embedded services on other websites.","_fbp","Used by Facebook to deliver a series of advertisement products such as real time bidding from third party advertisers","_zitok","_twpid","This cookie is typically associated with the use of Taptap Digital, a company specializing in digital advertising. It is used for tracking and ad personalization purposes.","_tt_enable_cookie","This cookie name is associated with TikTok which is used to collect data about behavior and purchases on the website and to measure the effect of advertising. It is also used to track website activity to help optimize advertising efforts.","AMCV_xxxxxAdobeOrg","This cookie is used to identify a unique visitor for tracking purposes by Adobe Experience Cloud.This is a pattern type cookie name associated with Adobe Marketing Cloud. It stores a unique visitor identifier, and uses an organisation identifier to allow a company to track users across their domains and services.","ttcsid","This cookie name is associated with TikTok which is used to enable TikTok Ads session tracking by assigning a unique session ID to users who click on TikTok ads, allowing for the attribution of user behavior and conversions within the same session.This cookie is associated with TikTok. It helps track your activity on the platform, including how you interact with their embedded services on other websites.","C0003","Personalization Cookies","These cookies collect data about how you have interacted with our website to help us improve your web experience, such as which pages you have visited. These cookies may store a unique ID so that our system will remember you when you return. They may be set by us or by third party providers whose services we have added to our pages. These cookies enable us to provide enhanced website functionality and personalization as well as make the marketing messages we send to you more relevant to your interests. If you do not allow these cookies, then some or all of these services may not function properly.",[590,591,592,593,594,595,596,597,598,599,600,601,602,603,604,605,606,607,608,609,610,611,612,613,614],[561,562,563,564,565],{"_399":475,"_243":475,"_210":401,"_372":585},{"_399":449,"_243":449,"_210":401,"_372":580},{"_399":575,"_243":575,"_210":401,"_372":576},{"_399":570,"_243":570,"_210":401,"_372":571},{"_399":566,"_243":566,"_210":401,"_372":567},"nvidia.partners",[568],{"_113":569,"_406":566,"_407":34,"_210":443},"edge","app.hushly.com",[572],{"_113":573,"_406":570,"_407":34,"_210":574},"*CORS","This domain is associated with Hushly, a platform that provides lead conversion solutions for content marketing by using personalization and content engagement strategies.","github.com",[577],{"_113":578,"_406":575,"_407":34,"_210":579},"_octo","This domain is owned by Github. The main business activity is: Source code repository and collaboration platform",[581,582],{"_113":584,"_406":449,"_407":34,"_210":453},{"_113":583,"_406":449,"_407":34,"_210":453},"lidc","li_gc",[586,587],{"_113":589,"_406":475,"_407":34,"_210":479},{"_113":588,"_406":475,"_407":34,"_210":479},"imid_*","imid",{"_113":615,"_406":524,"_407":34,"_210":617},{"_113":646,"_406":521,"_407":34,"_210":647},{"_113":615,"_406":645,"_407":34,"_210":617},{"_113":643,"_406":521,"_407":91,"_210":644},{"_113":641,"_406":521,"_407":34,"_210":642},{"_113":640,"_406":623,"_407":91,"_210":401},{"_113":615,"_406":619,"_407":34,"_210":617},{"_113":638,"_406":521,"_407":34,"_210":639},{"_113":615,"_406":527,"_407":34,"_210":617},{"_113":626,"_406":524,"_407":34,"_210":401},{"_113":615,"_406":637,"_407":34,"_210":617},{"_113":635,"_406":521,"_407":34,"_210":636},{"_113":632,"_406":633,"_407":34,"_210":634},{"_113":630,"_406":631,"_407":34,"_210":401},{"_113":615,"_406":627,"_407":34,"_210":617},{"_113":629,"_406":521,"_407":34,"_210":401},{"_113":615,"_406":628,"_407":34,"_210":617},{"_113":626,"_406":627,"_407":34,"_210":401},{"_113":625,"_406":623,"_407":91,"_210":401},{"_113":615,"_406":624,"_407":34,"_210":617},{"_113":615,"_406":623,"_407":34,"_210":617},{"_113":615,"_406":622,"_407":34,"_210":617},{"_113":621,"_406":521,"_407":91,"_210":617},{"_113":618,"_406":619,"_407":91,"_210":620},{"_113":615,"_406":616,"_407":34,"_210":617},"_gd_*","info.nvidia.com","This cookie contains no identifying information and is used for diagnostic purposes by Impact Radius, a service we use to track signup commissions for our affiliate program. We have categorized it under Functionality because it is necessary for the operation of our affiliate program.","nv_country_code","developer.nvidia.com","This cookie is likely used to store the country code of the visitor to provide localized content and possibly for legal compliance with regional data privacy laws.","_gdxxxxxxxxxxxxx","nvidianews.nvidia.com","docs.api.nvidia.com","research.nvidia.com","readme_library","wcmmode","ai-podcast.nvidia.com","resources.nvidia.com","geo_locator","openedx-language-preference","learn.nvidia.com","doxygen_width","raytracing-docs.nvidia.com","This cookie is likely used to remember the width preferences of a sidebar or interface element within a documentation platform like Doxygen.","_sfid_eb78","This cookie is categorized as Functionality Cookie.","crowdsource.nvidia.com","mbox","Cookie set by Adobe SiteCatalyst analytics software - used for measuring the performance of page content using A/B split testing","readme_language","vid","This cookie name 'vid' is used for tracking the search request made by the user on the 'Find a practitioner' page.","at_check","Used by Adobe Target to check if cookies are enabled/supported on the browser","blogs.nvidia.com","nvHeadless","This cookie is likely used to track if a user is using the site in a 'headless' non-GUI mode, which could be for accessibility reasons or for automated processes.","C0002","Performance Cookies","These cookies are used to provide quantitative measures of our website visitors, such as the number of times you visit, time on page, your mouse movements, scrolling, clicks and keystroke activity on the websites; other browsing, search, or product research behavior; and what brought you to our site. These cookies may store a unique ID so that our system will remember you when you return. Information collected with these cookies is used to measure and find ways to improve website performance.",[673,674,675,676,677,678,679,680,681,682,683,684,685,686,687,688,689,690,691,692,693,694,695,696,697,698,699,700,701],[653,654,655,656],{"_399":668,"_243":668,"_210":401,"_372":669},{"_399":663,"_243":663,"_210":401,"_372":664},{"_399":462,"_243":462,"_210":401,"_372":660},{"_399":575,"_243":575,"_210":401,"_372":657},[658],{"_113":659,"_406":575,"_407":91,"_210":579},"_gh_sess",[661],{"_113":662,"_406":462,"_407":91,"_210":469},"YSC","bizible.com",[665],{"_113":666,"_406":663,"_407":34,"_210":667},"_BUID","Bizible is a marketing attribution software that tracks the performance of marketing efforts and connects them to revenue, primarily used for B2B environments.","a.usbrowserspeed.com",[670],{"_113":671,"_406":668,"_407":34,"_210":672},"tuid","This domain is owned by Ookla which provides internet speed testing services through its Speedtest application.",{"_113":743,"_406":521,"_407":34,"_210":744},{"_113":728,"_406":742,"_407":91,"_210":729},{"_113":711,"_406":524,"_407":34,"_210":636},{"_113":711,"_406":624,"_407":34,"_210":636},{"_113":740,"_406":521,"_407":34,"_210":741},{"_113":728,"_406":527,"_407":91,"_210":729},{"_113":738,"_406":521,"_407":34,"_210":739},{"_113":728,"_406":524,"_407":91,"_210":729},{"_113":736,"_406":623,"_407":34,"_210":737},{"_113":728,"_406":645,"_407":91,"_210":729},{"_113":708,"_406":524,"_407":34,"_210":710},{"_113":734,"_406":521,"_407":91,"_210":735},{"_113":732,"_406":521,"_407":34,"_210":733},{"_113":730,"_406":521,"_407":91,"_210":731},{"_113":728,"_406":619,"_407":91,"_210":729},{"_113":726,"_406":521,"_407":34,"_210":727},{"_113":724,"_406":521,"_407":34,"_210":725},{"_113":722,"_406":521,"_407":91,"_210":723},{"_113":720,"_406":628,"_407":34,"_210":721},{"_113":708,"_406":619,"_407":34,"_210":710},{"_113":718,"_406":521,"_407":34,"_210":719},{"_113":716,"_406":521,"_407":91,"_210":717},{"_113":714,"_406":521,"_407":34,"_210":715},{"_113":712,"_406":521,"_407":34,"_210":713},{"_113":711,"_406":521,"_407":34,"_210":636},{"_113":708,"_406":709,"_407":34,"_210":710},{"_113":706,"_406":521,"_407":34,"_210":707},{"_113":704,"_406":521,"_407":91,"_210":705},{"_113":702,"_406":521,"_407":34,"_210":703},"_ga_xxxxxxxxxx","Used by Google Analytics to identify and track an individual session with your device.","s_ppvl","This cookie is set by Adobe Analytics to store information on the percentage of the page displayed to the end user.","coveo_visitorId","This cookie name is associated with Coveo which is used to store a unique user ID for marketing purposes and to track the number of distinct users who submitted a query or clicked an item, counting them only once even if they did the same query or clicked the same item over the course of many visits.","_dd_s","marketplace.nvidia.com","Datadog Cookie used to group all events generated from a unique user session across multiple pages. It contains the current session ID, whether the session is excluded due to sampling, and the expiration date of the session. The cookie is extended for an extra 15 minutes every time the user interacts with the website, up to the maximum user session duration (4 hours).","nvweb_A","_ga","This cookie name is associated with Google Universal Analytics - which is a significant update to Google's more commonly used analytics service. This cookie is used to distinguish unique users by assigning a randomly generated number as a client identifier. It is included in each page request in a site and used to calculate visitor, session and campaign data for the sites analytics reports. By default it is set to expire after 2 years, although this is customisable by website owners.","_uetvid","This is a cookie utilised by Microsoft Bing Ads and is a tracking cookie. It allows us to engage with a user that has previously visited our website.","tp","This cookie is associated with Adobe Analytics site tracking. It registers if the website's scroll-depth detection is active. This function remembers how far the user has scrolled on the current session across the website's sub-pages.","_hly_vid","This cookie _hly_vid is used to record website and server performance with each visit to the site. The data is aggregated and analyzed; therefore, it remains anonymous upon use.","_session_id","This cookie creates an interim session ID used as an in-session user ID.\n","s_cc","This cookie is associated with the Adobe Site Catalyst. It determines whether cookies are enabled in the web browser.","s_ecid","Adobe Analytics cookie containing the the Adobe Experience Cloud 1st party visitor id if the Analytics AMCV visitor id is not found.","_hjSessionUser_xxxxxx","Hotjar cookie that is set when a user first lands on a page with the Hotjar script. It is used to persist the Hotjar User ID, unique to that site on the browser. This ensures that behavior in subsequent visits to the same site will be attributed to the same user ID.","_hjHasCachedUserAttributes","This cookie name is associated with Hotjar which is used to know whether the data set in _hjUserAttributes Local Storage item is up to date or not.","s_ppv","This cookie is associated with Adobe Analytics site tracking and is used to measure a user's scroll activity to see how much of a page they view before moving to another page.","_gid","This cookie name is associated with Google Universal Analytics. This appears to be a new cookie and as of Spring 2017 no information is available from Google. It appears to store and update a unique value for each page visited.","_hjTLDTest","When the Hotjar script executes we try to determine the most generic cookie path we should use, instead of the page hostname. This is done so that cookies can be shared across subdomains (where applicable). To determine this, we try to store the _hjTLDTest cookie for different URL substring alternatives until it fails. After this check, the cookie is removed.","ekfls-2025-03-27","This cookie name is associated with readme.io which is used to help maintain data for the API reference support page.","s_getNewRepeat","This cookie is used to determine whether a visitor is a new visitor or a repeat visitor to the site, and is used for analytical purposes.","_hly_sid","This cookie is likely associated with website analytics, as the prefix '_hly_' suggests a connection to Holistics, a business intelligence and analytics platform.","docs.nvidia.com","_hjSession_xxxxxx","A cookie that holds the current session data. This ensues that subsequent requests within the session window will be attributed to the same Hotjar session.","C0001","Required Cookies","These cookies enable core functionality such as security, network management, and accessibility. These cookies are required for the site to function and cannot be turned off.",[890,891,892,893,894,895,896,897,898,899,900,901,902,903,904,905,906,907,908,909,910,911,912,913,914,915,916,917,918,919,920,921,922,923,924,925,926,927,928,929,930,931,932,933,934,935,936,937,938,939,940,941,942,943,944,945,946,947,948,949,950,951,952,953,954,955,956,957,958,959,960,961,962],[750,751,752,753,754,755,756,757,758,759,760,761,762,763,764,765,766,767,768,769,770,771,772,773,774],{"_399":887,"_243":887,"_210":401,"_372":888},{"_399":881,"_243":881,"_210":401,"_372":882},{"_399":878,"_243":878,"_210":401,"_372":879},{"_399":866,"_243":866,"_210":401,"_372":867},{"_399":863,"_243":863,"_210":401,"_372":864},{"_399":860,"_243":860,"_210":401,"_372":861},{"_399":855,"_243":855,"_210":401,"_372":856},{"_399":462,"_243":462,"_210":401,"_372":852},{"_399":849,"_243":849,"_210":401,"_372":850},{"_399":454,"_243":454,"_210":401,"_372":847},{"_399":449,"_243":449,"_210":401,"_372":845},{"_399":840,"_243":840,"_210":401,"_372":841},{"_399":835,"_243":835,"_210":401,"_372":836},{"_399":830,"_243":830,"_210":401,"_372":831},{"_399":827,"_243":827,"_210":401,"_372":828},{"_399":575,"_243":575,"_210":401,"_372":824},{"_399":434,"_243":434,"_210":401,"_372":821},{"_399":570,"_243":570,"_210":401,"_372":814},{"_399":566,"_243":566,"_210":401,"_372":808},{"_399":802,"_243":802,"_210":401,"_372":803},{"_399":797,"_243":797,"_210":401,"_372":798},{"_399":792,"_243":792,"_210":401,"_372":793},{"_399":788,"_243":788,"_210":401,"_372":789},{"_399":783,"_243":783,"_210":401,"_372":784},{"_399":775,"_243":775,"_210":401,"_372":776},"polyfill.io",[777,778],{"_113":781,"_406":775,"_407":34,"_210":782},{"_113":779,"_406":775,"_407":34,"_210":780},"cf_ob_info","This cookie is associated with sites using CloudFlare. It is used to improve page load times and to override any security restrictions based on the visitors IP address. It does not contain any user identification information.","cf_use_ob","This cookie is associated with sites using CloudFlare. It is used to improve page load times and to override any security restrictions based on the visitors IP address. It does not contain any user identification information. ","www.recaptcha.net",[785],{"_113":786,"_406":783,"_407":34,"_210":787},"_GRECAPTCHA","This domain is used by Google reCAPTCHA, which provides spam protection services to websites by detecting if a user is a human or a bot.","api2.hcaptcha.com",[790],{"_113":791,"_406":788,"_407":34,"_210":443},"__cflb","instagram.com",[794],{"_113":795,"_406":792,"_407":34,"_210":796},"csrftoken","Instagram is a social networking service owned by Facebook, Inc. It's designed for sharing photos and videos from a smartphone. The service also uses cookies for various purposes, including analytics, personalization, and advertising.","readme.io",[799],{"_113":800,"_406":797,"_407":34,"_210":801},"__cf_bm","The __cf_bm cookie is a cookie necessary to support Cloudflare Bot Management, currently in private beta. As part of our bot management service, this cookie helps manage incoming traffic that matches criteria associated with bots.","nvidiacorp.us-5.evergage.com",[804,805],{"_113":807,"_406":802,"_407":34,"_210":443},{"_113":806,"_406":802,"_407":34,"_210":443},"AWSALBTG","AWSALBTGCORS",[809,810],{"_113":812,"_406":566,"_407":34,"_210":813},{"_113":811,"_406":566,"_407":34,"_210":443},"ak_bmsc","bm_mi","This domain is likely associated with NVIDIA's partner program, which could involve marketing, sales, and support services for partners.",[815,816,817],{"_113":820,"_406":570,"_407":34,"_210":443},{"_113":819,"_406":570,"_407":91,"_210":443},{"_113":818,"_406":570,"_407":34,"_210":443},"AWSALBCORS","JSESSIONID","AWSALB",[822],{"_113":800,"_406":434,"_407":34,"_210":823},"This is a CloudFoundry cookie",[825],{"_113":826,"_406":575,"_407":34,"_210":579},"logged_in","hcaptcha.com",[829],{"_113":800,"_406":827,"_407":34,"_210":823},"tracking.intentsify.io",[832],{"_113":833,"_406":830,"_407":34,"_210":834},"userId","Used to track usage of the embedded video player.","static.reo.dev",[837],{"_113":838,"_406":835,"_407":34,"_210":839},"cookietest","Common cookie name could have a number of different origins. Where this is first party and a session cookie, its most likely to do with checking to see if the browser is set to block or allow cookies.","www.datadoghq-browser-agent.com",[842],{"_113":843,"_406":840,"_407":34,"_210":844},"dd_cookie_test_","dd_cookie_test",[846],{"_113":800,"_406":449,"_407":34,"_210":823},[848],{"_113":800,"_406":454,"_407":34,"_210":823},"supabase.co",[851],{"_113":800,"_406":849,"_407":34,"_210":823},[853],{"_113":854,"_406":462,"_407":34,"_210":469},"VISITOR_PRIVACY_METADATA","dosa-prd.azurewebsites.net",[857],{"_113":858,"_406":855,"_407":34,"_210":859},"access_token","This cookie is used to obtain the permissions for users of the portal to allow for API access in order to retrieve information required to perform portal functions.","w.hcaptcha.com",[862],{"_113":800,"_406":860,"_407":34,"_210":823},"cdn.readme.io",[865],{"_113":800,"_406":863,"_407":34,"_210":823},"queue-it.net",[868,869,870,871,872],{"_113":877,"_406":866,"_407":34,"_210":401},{"_113":876,"_406":866,"_407":34,"_210":401},{"_113":875,"_406":866,"_407":34,"_210":401},{"_113":874,"_406":866,"_407":34,"_210":401},{"_113":873,"_406":866,"_407":34,"_210":401},"Queue-it-{queueId}","Queue-it-{dmonvidia }________{eventId}","Queue-it-token","queueitsoftblock_{queueId}","Queue-it-{nvidia}________{eventId}","x.com",[880],{"_113":800,"_406":878,"_407":34,"_210":801},"api.hcaptcha.com",[883,884],{"_113":886,"_406":881,"_407":34,"_210":443},{"_113":791,"_406":881,"_407":34,"_210":885},"This domain is associated with hCaptcha, a service that provides CAPTCHA and reCAPTCHA alternatives for user verification and bot protection.","hmt_id","vimeo.com",[889],{"_113":800,"_406":887,"_407":34,"_210":823},{"_113":1023,"_406":521,"_407":34,"_210":1054},{"_113":1052,"_406":616,"_407":91,"_210":1053},{"_113":1045,"_406":1014,"_407":91,"_210":1046},{"_113":1051,"_406":709,"_407":91,"_210":401},{"_113":1049,"_406":521,"_407":34,"_210":1050},{"_113":807,"_406":1035,"_407":34,"_210":974},{"_113":982,"_406":1044,"_407":91,"_210":984},{"_113":806,"_406":973,"_407":34,"_210":974},{"_113":985,"_406":1044,"_407":34,"_210":986},{"_113":1047,"_406":521,"_407":91,"_210":1048},{"_113":1045,"_406":1021,"_407":91,"_210":1046},{"_113":979,"_406":1044,"_407":34,"_210":981},{"_113":1042,"_406":521,"_407":34,"_210":1043},{"_113":806,"_406":965,"_407":34,"_210":974},{"_113":970,"_406":1041,"_407":34,"_210":972},{"_113":818,"_406":971,"_407":34,"_210":987},{"_113":820,"_406":971,"_407":34,"_210":1038},{"_113":985,"_406":980,"_407":34,"_210":986},{"_113":1039,"_406":619,"_407":34,"_210":1040},{"_113":820,"_406":645,"_407":34,"_210":1038},{"_113":1036,"_406":521,"_407":34,"_210":1037},{"_113":812,"_406":645,"_407":34,"_210":1031},{"_113":1008,"_406":624,"_407":91,"_210":401},{"_113":806,"_406":1035,"_407":34,"_210":974},{"_113":1033,"_406":973,"_407":91,"_210":1034},{"_113":997,"_406":521,"_407":34,"_210":988},{"_113":970,"_406":521,"_407":34,"_210":972},{"_113":1032,"_406":533,"_407":34,"_210":401},{"_113":812,"_406":521,"_407":34,"_210":1031},{"_113":820,"_406":965,"_407":34,"_210":1030},{"_113":1028,"_406":628,"_407":91,"_210":1029},{"_113":1026,"_406":521,"_407":91,"_210":1027},{"_113":1025,"_406":521,"_407":34,"_210":401},{"_113":1023,"_406":971,"_407":34,"_210":1024},{"_113":1022,"_406":533,"_407":34,"_210":401},{"_113":1013,"_406":1021,"_407":91,"_210":1015},{"_113":1019,"_406":624,"_407":34,"_210":1020},{"_113":1017,"_406":709,"_407":34,"_210":1018},{"_113":1016,"_406":527,"_407":34,"_210":401},{"_113":1013,"_406":1014,"_407":91,"_210":1015},{"_113":1011,"_406":980,"_407":91,"_210":1012},{"_113":1009,"_406":521,"_407":34,"_210":1010},{"_113":1008,"_406":524,"_407":91,"_210":401},{"_113":979,"_406":983,"_407":34,"_210":981},{"_113":811,"_406":998,"_407":34,"_210":988},{"_113":1006,"_406":524,"_407":34,"_210":1007},{"_113":1004,"_406":521,"_407":34,"_210":1005},{"_113":1002,"_406":521,"_407":34,"_210":1003},{"_113":1001,"_406":521,"_407":34,"_210":401},{"_113":843,"_406":524,"_407":34,"_210":1000},{"_113":997,"_406":645,"_407":34,"_210":988},{"_113":800,"_406":616,"_407":34,"_210":999},{"_113":997,"_406":998,"_407":34,"_210":988},{"_113":996,"_406":533,"_407":34,"_210":401},{"_113":995,"_406":521,"_407":34,"_210":968},{"_113":993,"_406":980,"_407":34,"_210":994},{"_113":806,"_406":991,"_407":34,"_210":992},{"_113":990,"_406":533,"_407":34,"_210":401},{"_113":989,"_406":521,"_407":34,"_210":401},{"_113":982,"_406":980,"_407":91,"_210":984},{"_113":811,"_406":645,"_407":34,"_210":988},{"_113":818,"_406":645,"_407":34,"_210":987},{"_113":985,"_406":983,"_407":34,"_210":986},{"_113":982,"_406":983,"_407":91,"_210":984},{"_113":979,"_406":980,"_407":34,"_210":981},{"_113":977,"_406":973,"_407":91,"_210":978},{"_113":975,"_406":631,"_407":34,"_210":976},{"_113":807,"_406":973,"_407":34,"_210":974},{"_113":970,"_406":971,"_407":34,"_210":972},{"_113":800,"_406":623,"_407":34,"_210":969},{"_113":967,"_406":521,"_407":34,"_210":968},{"_113":818,"_406":965,"_407":34,"_210":966},{"_113":963,"_406":521,"_407":34,"_210":964},"LoginHints","Keeps users logged in across NVIDIA domains.","api.ngc.nvidia.com","This cookie is managed by AWS and is used for load balancing.","bm_so","This cookie name is used to load the site as quickly as possible and to recognize bots using Akamai.","The __cf_bm cookie is a cookie necessary to support Cloudflare Bot Management, currently in private beta. As part of our bot management service, this cookie helps manage incoming traffic that matches criteria associated with bots.\nThis is a CloudFoundry cookie","OptanonConsent","nvid.nvidia.com","This cookie is set by the cookie compliance solution from OneTrust. It stores information about the categories of cookies the site uses and whether visitors have given or withdrawn consent for the use of each category. This enables site owners to prevent cookies in each category from being set in the users browser, when consent is not given. The cookie has a normal lifespan of one year, so that returning visitors to the site will have their preferences remembered. It contains no information that can identify the site visitor.","events.nvidia.com","This cookie is associated with the Amazon Web Services Elastic Load Balancing functionality for routing client request on the server.Used to attribute commission to affiliates when you arrive at the website from an affiliate referral link. It is set when you click on one of our links and used to let the advertiser and us know the website from which you came.","sessionid","This is a very generic cookie name that may have different purposes on different sites, but generally it will be some kind of anonymous session identifier.","_csrf","This cookie is used to prevent Cross-Site Request Forgery (CSRF) attacks, ensuring that each request made to the server is from the user's browser and is intentional.","LSKey-c$CookieConsentPolicy","enterprise-support.nvidia.com","Used to apply end-user cookie consent preferences set by our client-side utility.Custom- Short Duration. Unique ID of an Enquiry- used to validate access to case, and load exisiting data.\n","renderCtx","academicgrants.nvidia.com","Used to store site parameters in the session for reuse across requests by a single client for functionality and performance reasons.","CookieConsentPolicy","Used to apply end-user cookie consent preferences set by our client-side utility.","This cookie is managed by AWS and is used for load balancing.Used to attribute commission to affiliates when you arrive at the website from an affiliate referral link. It is set when you click on one of our links and used to let the advertiser and us know the website from which you came.","This cookie is associated with Akamai and is used to differentiate between traffic from humans and bots.","nv_co","__uxq412__ses.88c1","learn.learn.nvidia.com","This cookie is associated with the Amazon Web Services Elastic Load Balancing functionality for routing client request on the server.","pctrk","Used to count page views by unauthenticated users against license usage.","bm_s","bpazaws52gukakzc__ctrl0_ctl36_uccaptcha","bm_sv","store.nvidia.com","The __cf_bm cookie is a cookie necessary to support Cloudflare Bot Management, currently in private beta. As part of our bot management service, this cookie helps manage incoming traffic that matches criteria associated with bots.\n","This cookie name is used by DataDog to determine if cookies are supported on the website.dd_cookie_test","AMP_TLDTEST_xxxxxxxx","SESSlsd","This cookie is likely to be a session identifier used to maintain user session state across page requests.","mboxEdgeCluster","This cookie name is associated with Adobe Target which is used to communicate the correct edge cluster to at.js so that Target profiles can stay in sync as users navigate across a site.","_rdt_em","This cookie name is associated with Reddit which is used to manage session state and user interactions.","megaNavCollapsed","_abck","The _abck cookie is associated with Akamai's Bot Manager, which is used to prevent fraud and abuse by safeguarding online interactions and detecting irregularities that could suggest malicious activity, such as bot-driven attacks or hacking attempts.","idccsrf","This cookie name is associated with Salesforce which is used to track CrossSiteRequestForgery validation for certain single sign-on (SSO) flows.","ApplicationGatewayAffinity","partners.nvidia.com","This cookie is used to maintain a user session by the server in a load balanced environment, to ensure the user has a consistent experience.","AMP_TEST_xxxxxxxx","bm_lso","This cookie name is associated with Akamai which is used to load the site as quickly as possible and to recognize bots.","SSESS###########","This cookie is likely to be a session cookie used to maintain an anonymized user session by the server in a secure environment, often seen in sites built with the Drupal content management system.","nvdeveloper.nvidia.com","__uxq412__id.88c1","OptanonAlertBoxClosed","This cookie is set by websites using certain versions of the cookie law compliance solution from OneTrust. It is set after visitors have seen a cookie information notice and in some cases only when they actively close the notice down. It enables the website not to show the message more than once to a user. The cookie has a normal lifespan of one year and contains no personal information.","QueueITAccepted-SDFrts345E-V3_{eventId}","TEST_AMCV_COOKIE_WRITE","This cookie is used to test whether cookies are enabled in the browser. It is a temporary cookie used to determine if the browser accepts cookies.","XSRF-TOKEN","This cookie is written to help with site security in preventing Cross-Site Request Forgery attacks.This cookie helps with site security in preventing Cross-Site Request Forgery attacks.","AWS ELB application load balancer","This cookie is associated with Akamai. It is used to analyse traffic to determine if it is automated traffic generated by bots or a human user.","bpazaws52gukakzc__ctrl0_ctl78_list_ctl09_systemcaptcha","PHPSESSID","PHP session cookie associated with embedded content from this domain.\n","base.manager.learn.nvidia.com","bm_ss","This cookie is used to help protect websites from automated bot traffic.","AWS ELB application load balancerUsed to attribute commission to affiliates when you arrive at the website from an affiliate referral link. It is set when you click on one of our links and used to let the advertiser and us know the website from which you came.","_devzone_session","This cookie is likely used to maintain a user session on the website by the server. It is common for session identifiers to be used to keep users logged in and track their interaction with the site during a session.","docs.ngc.nvidia.com","bm_sz","This cookie is used by Akamai to optimize site performance and security. It is typically associated with the delivery of web content and services.","vcalliance.nvidia.com","ASP.NET_SessionId","General purpose platform session cookie, used by sites written with Microsoft .NET based technologies. Usually used to maintain an anonymised user session by the server.","com.adobe.alloy.getTld","This cookie name is associated with the Adobe WebSDK which is used to determine the top-most domain that can accept cookies.","SESSls","This cookie is likely to be a session identifier, used to maintain a user's session state across page requests.","c_code","BIGipServerx","This cookie name is associated with the BIG-IP product suite from company F5. Usually associated with managing sessions on load balanced servers, to ensure user requests are routed consistently to the correct server. The common root is BIGipServer most commonly followed by a domain name, usually the one that it is hosted on, but not always.","This cookie is set by websites using certain versions of the cookie law compliance solution from OneTrust. It is set after visitors have seen a cookie information notice and in some cases only when they actively close the notice down. It enables the website not to show the message more than once to a user. The cookie has a one year lifespan and contains no personal information.","default","

NVIDIA uses cookies to improve your experience on our web site. We and our third-party partners also use cookies and other tools to collect and record information you provide as well as information about your interactions with our websites for performance improvement, analytics, and to assist in marketing efforts. By clicking \"Accept All\", you consent to our use of cookies and other tools as described in our Cookie Policy. You can manage your cookie settings by clicking on \"Manage Settings.\" By continuing to use this site or by clicking one of the buttons below, you agree to our Terms of Service (which contains important waivers). Please see our Privacy Policy for more information on our privacy practices.

","gpc","

We have detected the Global Privacy Control (GPC) signal and recorded your rejection of all optional cookies on this site for this browser. You can manage your cookie settings by clicking on \"Manage Settings\". Please see our Cookie Policy for more information. To opt out of non-cookie personal information \"sales\" / \"sharing\" for targeted advertising purposes, please visit the NVIDIA Preference Center. Please see our Privacy Policy for more information on our privacy practices.

","gpcLoggedIn","

We have detected the Global Privacy Control Signal (GPC) and have opted you out of all optional cookies on this browser. You can manage your cookie settings by clicking on \"Manage Settings\". Please see our Cookie Policy for more information. We have also opted you out of \"sharing\"/\"sales\" of personal information outside of cookies. You can manage these settings in the NVIDIA NVIDIA Preference Center. Please see our Privacy Policy for more information.

","gpcLoggedInOverridden","

We have detected the Global Privacy Control Signal (GPC) and have opted you out of all optional cookies on this browser. You can manage your cookie settings by clicking on \"Manage Settings\". Please see our Cookie Policy for more information. We have also opted you out of \"sharing\"/\"sales\" of personal information outside of cookies which overrides at least one of your previous settings. You can manage them in the NVIDIA Preference Center. Please see our Privacy Policy for more information.

",{"_745":16,"_648":59,"_556":59,"_376":59,"_366":59},"hasConsented","gpcApplied","showBanner","geolocation",{"_1069":1070,"_1071":1072},"country","US","state","OR","query","categories",[1092,1093,1094,1095,1096],"optionUrls",[1080,1081,1082,1083,1084,1085,1086,1087,1088,1089,1090,1091],"total",12,"/orgs/nim/deepseek-ai/containers/deepseek-r1/-","/orgs/nim/nvidia/containers/llama-3.1-nemotron-70b-instruct/-","/orgs/nvidia/-/containers/pytorch/-","/orgs/nvidia/omniverse/collections/kit/-","/orgs/nvidia/-/collections/deepstream_sdk/-","/orgs/nvidia/omniverse/collections/kit-appstreaming-collection/-","/orgs/nvidia/research/models/stylegan3/-","/orgs/nvidia/tao/models/peoplenet/-","/orgs/nvidia/tao/models/trafficcamnet/-","/orgs/nvidia/riva/resources/riva_quickstart/-","/orgs/nvidia/-/helm-charts/gpu-operator/-","/orgs/nvidia/-/helm-charts/k8s-nim-operator/-",{"_20":1170,"_1098":1171,"_1100":59},{"_20":1135,"_1098":1136,"_1100":57},{"_20":1118,"_1098":1119,"_1100":1120},{"_20":21,"_1098":1111,"_1100":1112},{"_20":1097,"_1098":1099,"_1100":19},"helm-charts","items",[1101,1102],"startIndex",{"_20":1097,"_243":1108,"_204":205,"_206":-7,"_238":1109,"_246":1105,"_271":1110,"_1107":1090},{"_20":1097,"_243":1103,"_204":205,"_206":-7,"_238":1104,"_246":1105,"_271":1106,"_1107":1091},"NVIDIA NIM Operator","k8s-nim-operator","NVIDIA",[],"url","GPU Operator","gpu-operator",[],[1113],9,{"_20":21,"_243":1114,"_204":205,"_206":1115,"_238":1116,"_246":1105,"_271":1117,"_1107":1089},"Riva Skills Quick Start","riva","riva_quickstart",[],"models",[1121,1122,1123],6,{"_20":1118,"_243":1131,"_204":205,"_206":1132,"_238":1133,"_246":1105,"_271":1134,"_1107":1086},{"_20":1118,"_243":1128,"_204":205,"_206":1125,"_238":1129,"_246":1105,"_271":1130,"_1107":1087},{"_20":1118,"_243":1124,"_204":205,"_206":1125,"_238":1126,"_246":1105,"_271":1127,"_1107":1088},"TrafficCamNet","tao","trafficcamnet",[],"PeopleNet","peoplenet",[],"StyleGAN3 pretrained models","research","stylegan3",[],"collections",[1137,1138,1139],{"_20":1135,"_243":1167,"_204":205,"_206":1141,"_238":1168,"_246":1105,"_271":1169,"_1107":1083},{"_20":1135,"_243":1164,"_204":205,"_206":-7,"_238":1165,"_246":1105,"_271":1166,"_1107":1084},{"_20":1135,"_243":1140,"_204":205,"_206":1141,"_238":1142,"_246":1105,"_271":1143,"_1107":1085},"Omniverse Kit App Streaming","omniverse","kit-appstreaming-collection",[1144,1145,1146],{"_113":1158,"_243":1159,"_240":1160,"_250":1161,"_1151":1162},{"_113":1141,"_243":1154,"_240":1155,"_250":1150,"_1151":1156},{"_113":1147,"_243":1148,"_240":1149,"_250":1150,"_1151":1152},"omniverse-dev","NVIDIA Developer Program","Join the Developer Program for access to free tools, support, and tech resources.","https://store-images.s-microsoft.com/image/apps.33936.6f1bcf42-55d0-45d1-b4b6-f08c68a50bbe.e91ca707-db64-49d6-9761-4781590a9685.7ba64fc5-7c44-4864-a5e6-7bde05126935","supportedFeatures",[1153],"UNIFIED_CATALOG","NVIDIA Omniverse","NVIDIA Omniverse is a platform of APIs, services, and software development kits (SDKs) that enable developers to build generative AI-enabled tools and applications.",[1153,1157],"SHAREABLE_TO_USER_GROUPS","nv-ai-enterprise","NVIDIA AI Enterprise","Accelerate your AI agent development","https://assets.ngc.nvidia.com/products/catalog/images/product-NVAIE.jpg",[1153,1157,1163],"NGC_CONSOLE","DeepStream SDK","deepstream_sdk",[],"Omniverse Kit (FB)","kit",[1144,1145,1146],"containers",[1172,1173,1174],{"_20":1170,"_243":1182,"_204":1179,"_206":1183,"_238":1184,"_246":1105,"_271":1185,"_1107":1080},{"_20":1170,"_243":1178,"_204":1179,"_206":205,"_238":1180,"_246":1105,"_271":1181,"_1107":1081},{"_20":1170,"_243":1175,"_204":205,"_206":-7,"_238":1176,"_246":1105,"_271":1177,"_1107":1082},"PyTorch","pytorch",[],"Llama-3.1-Nemotron-70B-Instruct","nim","llama-3.1-nemotron-70b-instruct",[],"DeepSeek-R1","deepseek-ai","deepseek-r1",[],"ownerOrgs",[],"memberOrgs",[],"externalOrgs",[],"apiKey","2b44f86fea50e90ffd2db29ee2028419","consentGranted","applicationId","4c857e4f-5ffa-45fa-b75e-af773c22f0ad","clientToken","pubb6927db33dae60ba5031a186cde0242a","site","datadoghq.com","service","catalog-ui","env","sessionSampleRate",100,"sessionReplaySampleRate","compressIntakeRequests","storeContextsAcrossPages","defaultPrivacyLevel","mask-user-input","trackViewsManually","trackingConsent","not-granted",{"_1215":1216,"_1217":313,"_1218":34,"_1219":1220,"_1098":1221,"_1222":1223,"_1224":1216,"_1225":34,"_1226":34,"_1227":34,"_1228":1229},"appName","ngc-catalog","appOrigin","hideOnPrivateOrg","icon","ObjectsModules2",[1230,1231,1232,1233,1234,1235,1236],"menuIcon","editor-swatches","productName","requiresAllScopes","requiresLogin","requiresOrgLevelScopesOnly","title","Catalog",{"_1228":1258,"_50":1259,"_1222":1260},{"_1251":1252,"_1228":1253,"_50":1254,"_1224":1158},{"_1228":1249,"_50":1250,"_1222":1223},{"_1228":1246,"_50":1247,"_1222":1248},{"_1228":1243,"_50":1244,"_1222":1245},{"_1228":1240,"_50":1241,"_1222":1242},{"_1228":1237,"_50":1238,"_1222":1239},"Resources","/resources","misc-library","Models","/models","misc-neural-network","Helm Charts","/helm-charts","misc-ship-wheel","Containers","/containers","3d-cube","Collections","/collections","access",[1255],"Enterprise Catalog","/enterprise",[1256,1257],"NVAIE","READ","Explore Catalog","/","common-home",{"_157":158,"_159":160,"_161":162,"_163":164,"_165":166,"_167":168},{"_156":34}]