MLCommons recently released the results of the MLPerf Training v6.1 benchmark. Two new benchmarks and the submission data fully demonstrate the latest architectural advances in AI training systems and the growing diversity of training hardware.
New Benchmarks Cover Recommendation and Large-Scale Post-Training
The MLPerf Training benchmark suite evaluates the combined performance of machine learning applications across models, software, and hardware through full-system testing. Version 6.1 adds the DLRMv4-HSTU recommendation benchmark and the Qwen3.5-397B large-scale LLM post-training benchmark.
DLRMv4-HSTU adopts an HSTU-based attention architecture and is trained on the Yambda dataset, replacing the previous DLRM-dcnv2. Qwen3.5-397B is the first large-scale post-training workload introduced by MLPerf; it uses Group Relative Policy Optimization (GRPO) and focuses on testing system reasoning capabilities on software engineering tasks such as tool calling.
Diversity of Submitted Systems Increases Significantly
For v6.1, 24 organizations submitted 153 systems, producing 250 test results. Of these, 37% were cloud systems, up from 29% in v6.0. Twelve systems used more than 1,000 accelerators, with the largest reaching 10,240 accelerators. Hardware covered products from NVIDIA, AMD, and Google, with NVIDIA Rubin GPUs and Google TPUv7 (Ironwood) appearing for the first time.
Industry Participation Reaches a New High
Submitting organizations include AMD, ASUSTeK, Azure, Cisco, CoreWeave, Crusoe, Dell, Fujitsu, Google, HPE, NVIDIA, Oracle, Supermicro, and others, with Crusoe among the first-time submitters. MLCommons said broad industry participation strengthens the AI ecosystem and continues to drive benchmarks toward scenarios that more closely reflect real-world use.
Full results and visualization tools are available on the MLCommons website.
© 2026 Winzheng.com 赢政天下 | 转载请注明来源并附原文链接