DLRM - AshokBhat/ml GitHub Wiki
About
- Deep learning recommendation model
- Introduced by Facebook research
- https://ai.facebook.com/blog/dlrm-an-advanced-open-source-deep-learning-recommendation-model/
DLRM v1 vs V2
| Feature/Aspect | DLRMv1 | DLRMv2 |
|---|---|---|
| Use Case Representation | More realistic large-scale recommender system benchmark | |
| Model Architecture | Uses an interactions layer to combine features | Replaces interactions layer with a three-layer DCNv2 cross network |
| Input Encoding | One-hot categorical inputs | Multi-hot categorical inputs |
| Dataset | Smaller, less complex dataset | Larger, more complex dataset |
| Hardware Utilization | Less demanding on memory bandwidth | Requires higher memory bandwidth |
MLPerf numbers (DLRMv2 99.99)
| Version | Hardware | Software | CPU | Cores | Offline(samples/s) | Server (samples/s) |
|---|---|---|---|---|---|---|
| 5.0 | 1 Node, 2 socket, GR | PyTorch | Xeon 6787P | 86 | 12397 | 11788 |
| 5.0 | 1 Node, 2 socket, GR | PyTorch | Xeon 6980P | 128 | 18686 | 18117 |
| 4.1 | 1 Node, 2 socket, EMR | PyTorch | Xeon 8592+ | 64 | 9830 | 9101 |
| 5.0 | GH200 1-GPU | TensorRT | Grace CPU | 51970 | 50070 | |
| 5.0 | L40S 8-GPU | TensorRT | Xeon 6740E | 100,517 | 94,989 |