GPUNet for PyTorch | NVIDIA NGC

NVIDIA Deep Learning Examples

GPUNet for PyTorch

Resource

NVIDIA Deep Learning Examples

GPUNet for PyTorch

GPUNet is a new family of Convolutional Neural Networks crafted by NVIDIA AI.

This section demonstrates the GPUNet training and inference results independently benchmarked by a third party. You can also easily replicate the same results following Quick Start Guide.

Results

Training Accuracy Results

We benchmark the training results following the steps in Training. This section lists the training results on NVIDIA DGX V100.

NVIDIA DGX V100 (8x V100 32GB)

Model	Batch	Epochs	GPUs	FP32 Top1	AMP Top1	FP32 (hours) Train Time	AMP (hours) Train Time	Training speedup (FP32 / AMP)
GPUNet-0	192	450	8	78.90_+/-0.03	78.96_+/-0.05	71.63	46.56	1.54 x
GPUNet-1	192	450	8	80.4-_+/-0.03	80.5_+/-0.03	67.5	43.5	1.55 x
GPUNet-2	192	450	8	82.1-_+/-0.04	82.2_+/-0.04	171	84.25	2.03 x

Training performance results

Please also follow the steps in Training to reproduce the performance results below.

NVIDIA DGX V100 (8x V100 32GB)

Model	GPUs	Batch	FP32 imgs/second	AMP imgs/second	Speedup (FP32 to AMP)
GPUNet-0	8	192	2289 img/s	3518 img/s	1.53 x
GPUNet-1	8	192	2415 img/s	3774 img/s	1.56 x
GPUNet-2	8	192	948 img/s	1957 img/s	2.03 x

NVIDIA DGX A100 (8x A100 80GB)

Model	GPUs	Batch	FP32 imgs/second	AMP imgs/second	Speedup (TF32 to AMP)
GPUNet-2	8	192	2002 img/s	2690 img/s	1.34 x
GPUNet-D1	8	128	755 img/s	844 img/s	1.11 x

Inference results

We benchmark the training results following the steps in Benchmark the GPUNet latency. This section lists the inference results on NVIDIA 32G V100 and 80G A100.

NVIDIA V100 (32GB)

GPUNet	Batch	GPU	TensorRT8 FP16 Latency	FP16 Latency	Perf Details	ImageNet Top1
GPUNet-0	1	V100	0.63 ms	1.82 ms	here	78.9
GPUNet-1	1	V100	0.82 ms	2.75 ms	here	80.5
GPUNet-2	1	V100	1.68 ms	5.50 ms	here	82.2
GPUNet-P0	1	V100	0.63 ms	2.11 ms	here	80.3
GPUNet-P1	1	V100	0.96 ms	2.47 ms	here	81.1
GPUNet-D1	1	V100	1.24 ms	2.88 ms	here	82.5
GPUNet-D2	1	V100	2.17 ms	4.22 ms	here	83.6

NVIDIA A100 (80GB)

GPUNet	Batch	GPU	TensorRT8 FP16 Latency	FP16 Latency	Perf Details	ImageNet Top1
GPUNet-0	1	A100	0.46 ms	1.46 ms	here	78.9
GPUNet-1	1	A100	0.59 ms	1.81 ms	here	80.5
GPUNet-2	1	A100	1.25 ms	4.03 ms	here	82.2
GPUNet-P0	1	A100	0.45 ms	1.31 ms	here	80.3
GPUNet-P1	1	A100	0.61 ms	1.64 ms	here	81.1
GPUNet-D1	1	A100	0.94 ms	2.44 ms	here	82.5
GPUNet-D2	1	A100	1.40 ms	3.06 ms	here	83.6