LLM Pretraining on MLPerf C4
20Train step throughputMXFP4 + H16
Evaluation Results
| Method | Links | |||
|---|---|---|---|---|
| MXFP4 + H16Backbone=Llama 3.1-8B, Hardware=MI355X GPUs, Numerical Format=MXFP4, Stabilization Technique=Deterministic Hadamard (H16)2026.05 | 20 | 8 | 9 |
Loading the SOTA2 catalog…