Release Intel® AI Reference Models v3.0.0 · intel/ai-reference-models

New features

Model Zoo for Intel® Architecture is rebranded as Intel® AI Reference Models to reflect it's purpose of showcasing to external audiences the internally achieved best performance configurations for critical workloads on Intel® Architecture.
Updates Intel® Data Center GPU Max Series 1550 x4 OAM workloads scripts for the best known configurations.
Introduces the initial version of Jupyter notebook interface for Intel® AI Reference Models.
Distributed training is supported for the following PyTorch models with different precisions: ResNet50, SSD-ResNet34, DLRM, MaskRCNN, RNNT and BERT Large.
Updated Transfer Learning Jupyter notebooks.
New supported workloads:
- MemRec DLRM FP32 inference: introducing Memory Efficient Recommendation System using Alternative Representation (MemRec). MemRec is a technology for alternative representation of embedding tables. It uses bloom filters and hashing techniques to encode embedding tables in a much smaller memory footprint optimized to make use of hierarchical cache architecture of Intel Xeon platforms. MemRec encodes DLRM embedding tables into two cache-friendly embedding tables to maximize predictive performance and increase recommendation accuracy.
- DLRM v2 training and inference with different precisions for GPU and CPU platforms.
This release contains many bug and CVE fixes to the previous versions.

Intel® AI Reference Models v3.0.0 is validated on the following environment: