New features of MLPerf Inference v2.1include:
- A new object detection benchmark: RetinaNet using a ResNeXT50 backbone. This reference model is more accurate than its predecessor and trains on the Open Images dataset, which is larger and more realistic than COCO.
- Inference over the network scenario: A remote system injects queries over the network into the SUT. This scenario more faithfully models real data center systems compared to starting with queries in host DRAM.