The NCF model focuses on providing recommendations. This is a modified implementation with improved overfitting and better accuracy.
Performance
Results
The following sections provide details on how we achieved our performance and accuracy in training and inference.
Training accuracy results
The following table lists the best hit rate at 10 for DGX-1 with 8 V100 32G GPUs:
| Number of GPUs | Full precision HR@10 | Mixed precision HR@10 |
|---|---|---|
| 1 | 0.959015 | 0.959485 |
| 4 | 0.959389 | 0.959274 |
| 8 | 0.959015 | 0.96 |
Here's an example validation accuracy curve for mixed precision vs full precision on DGX-1 with 8 V100 32G GPUs:
The histogram below shows the best HR@10 achieved
for 400 experiments using mixed precision and 400 experiments using single precision.
Mean HR@10 for mixed precision was equal to 0.95917 and for single precision it was equal to
0.95915.
Training performance results
NVIDIA DGX-1 with 8 V100 16G GPUs
The following table shows the best training throughput:
| Number of GPUs (samples/sec) | Mixed precision (samples/sec) | Full precision (samples/sec) | Speedup |
|---|---|---|---|
| 1 | 20,027 840 | 9,529,271 | 2.10 |
| 4 | 62,633 260 | 32,719,700 | 1.91 |
| 8 | 99,332 230 | 55,004,590 | 1.81 |
NVIDIA DGX-1 with 8 V100 32G GPUs
The following table shows the best training throughput:
| Number of GPUs (samples/sec) | Mixed precision (samples/sec) | Full precision (samples/sec) | Speedup |
|---|---|---|---|
| 1 | 18,871 650 | 9,206 424 | 2.05 |
| 4 | 59,413 640 | 31,898 870 | 1.86 |
| 8 | 94,752 770 | 53,645 640 | 1.77 |
Inference performance results
NVIDIA DGX-1 with 8 V100 16G GPUs
The following table shows the best inference throughput:
| Number of GPUs (samples/sec) | Mixed precision (samples/sec) | Full precision (samples/sec) | Speedup |
|---|---|---|---|
| 1 | 58,836,420 | 28,964,964 | 2.03 |
NVIDIA DGX-1 with 8 V100 32G GPUs
The following table shows the best inference throughput:
| Number of GPUs (samples/sec) | Mixed precision (samples/sec) | Full precision (samples/sec) | Speedup |
|---|---|---|---|
| 1 | 55,317,010 | 28,470,920 | 1.94 |