MLPerf HPC — AI Training on Supercomputers
MLPerf HPC measures the time to train specific scientific AI models to a target accuracy on HPC systems. Unlike general MLPerf benchmarks focused on cloud AI, MLPerf HPC uses scientific datasets and evaluates scaling efficiency across many nodes — the definitive standard for AI-capable HPC procurement.
CosmoFlow
3D convolutional neural network trained on N-body cosmology simulations to predict universe structure parameters. Representative of scientific CNN workloads with large 3D volumetric input data — common in climate science, medical imaging, and geophysics.
DeepCAM
Segmentation network trained on climate model output to detect extreme weather events (atmospheric rivers, tropical cyclones). Exercises image segmentation at scale — represents the largest class of scientific AI workloads in terms of dataset size.
OpenCatalyst
Graph neural network predicting energy of molecular systems — critical for computational chemistry and materials discovery. Exercises sparse, irregular compute patterns very different from dense tensor operations in standard AI benchmarks.
BERT Pre-training
Training BERT (Bidirectional Encoder Representations from Transformers) to target masked language modeling accuracy. Tests large-scale transformer training — the dominant architecture for LLMs — on HPC interconnects with hundreds to thousands of GPUs.
Time-to-Train Metric
MLPerf measures time from data loading start to reaching a specified target accuracy — not throughput or FLOPS. This end-to-end metric includes data preprocessing, communication overhead, and I/O latency, making it a realistic measure of actual AI productivity on HPC systems.
Scaling Efficiency
MLPerf HPC results are reported at multiple node counts, revealing scaling efficiency. Systems with fast interconnects (InfiniBand NDR, NVLink) maintain high efficiency as scale increases. Poor scaling indicates communication bottlenecks that will impact all large AI training jobs.
Biannual Updates
MLCommons publishes new MLPerf HPC results biannually, adding new benchmarks as scientific AI evolves. This ensures procurement decisions reflect current AI workloads — unlike HPL which has been unchanged since 1993. Always reference the most recent round when comparing systems.
Using MLPerf for Procurement
Request MLPerf HPC results on the specific GPU hardware you are procuring. Compare time-to-train across vendors at your target node count. Verify that the fabric (InfiniBand vs Ethernet) does not degrade scaling beyond 64 nodes — this is where the interconnect becomes the bottleneck.