MLPerf is a set of benchmarks designed to measure the performance of a machine learning model on a target system. Each benchmark measures the wall time to train a model to a target quality metic.
MLPerf is a trademark and is maintained by MLCommons and as such if publishing results of these benchmarks in a public work, use their guidelines.
MLPerf Training (v2.1)
Benchmark | Model | Dataset | Domain Area | Benchmark Target Metric | NCSA Notes | NCSA Results (If Available) |
---|---|---|---|---|---|---|
Image Classification | ResNet | ImageNet | Vision | 75.90% classification | Doesn't have download and prep steps commited to repo | |
Image Segmentation | 3D U-Net | KiTS19 | Vision | 0.908 Mean DICE score | Had issues with apptainer conversion | |
Natural Language Processing | BERT | Wikipedia | Language | 0.72 Mask-LM accuracy | ||
Object Detection (light-weight) | RetinaNet | OpenImages | Vision | 34.0% mAP | ||
Object Detection (heavy-weight) | Mask R-CNN | COCO | Vision | 0.377 Box min AP and 0.339 Mask min AP | ||
Recommendation | DLRM | 1TB Clickthrough | Commerce | 0.8025 AUC | ||
Reinforcement learning | Minigo | GO | Research | 50% win rate vs. checkpoint | ||
Speech Recognition | RNN-T | LibriSpeech | Language | 0.058 Word Error Rate |
MLPerf Inference (v2.1)
Benchmark | Dataset | Model | Domain benchmark represents | NCSA Notes | NCSA Results (If Available) |
---|---|---|---|---|---|
Image Classification | ImageNet2012 | ResNet50 | |||
Image Classification | OpenImages | ResNext50 | |||
Image Segmentation | KiTS19 | 3D U-Net | |||
Natural Language Processing | Squad 1.1 | BERT | |||
Recommendation | Criteo Terabyte | DLRM | |||
Speech Recognition | OpenSLR LibriSpeech Corpus | RNN-T |
0 Comments