This is a clone of the official MLCommons repository providing the most recent MLPerf benchmarks and results. The Juice benchmarks in closed/Juice are based on the NVIDIA-optimized benchmarks in closed/NVIDIA with changes to run with Juice. The changes amount to installing and using Juice to dynamically attach GPUs in the Docker container for running benchmarks.
To run a benchmark using Juice copy the Juice archive juice-gpu-linux.tar.gz for the Juice build you want to benchmark to closed/Juice/docker/common then follow the same instructions for running the NVIDIA benchmarks but running commands from the closed/Juice directory instead of closed/NVIDIA. The steps in setting up and running benchmarks are the same up until the final two steps of adding a custom system and running the benchmark -- launch these steps with juice run to attach a Juice GPU dynamically. For example the last two steps of running the Resnet50 benchmark and Offline scenario with Juice:
# ...
juice login
juice run python3 -m scripts.custom_systems.add_custom_system
juice run make run RUN_ARGS="--benchmarks resnet50 --scenarios offline"See Running Juice MLPerf Inference v5.0 Benchmarks for more details.