Nvidia has given its rivals another headache after dominating the latest MLPerf Training v6.1 benchmarks, which tested 153 computing systems across 250 results.
MLCommons tests attracted submissions from 24 organisations, including AMD, Google, Microsoft and Nvidia, measuring how quickly systems train AI models to specified accuracy targets.
Nvidia claims its new Vera Rubin NVL72 platform delivered up to twice the performance per GPU of its previous GB300 NVL72 system in selected comparisons. The company demonstrated impressive scaling on the DeepSeek-V3 benchmark, reducing training time from approximately 16 minutes to four by increasing its GPU count from 256 to 1,024.
Nvidia submitted results using 10,240 GB300 accelerators, while Microsoft’s Azure operation matched that scale with Nvidia GB200 hardware.
AMD made its presence felt through cloud provider Crusoe, which tested Instinct MI355X accelerators using standard Ethernet networking. Crusoe’s 128-GPU configuration completed the Llama 3.1 8B training benchmark in 14.07 minutes, compared with 78.34 minutes using eight GPUs. Its 64-GPU system completed the new DLRMv4 recommendation benchmark in 4.63 minutes, demonstrating that AMD hardware can scale effectively without Nvidia’s proprietary networking technology.
Google participated with its TPU processors, including its latest Ironwood generation, although Nvidia remains the company everyone else is trying to catch. MLCommons introduced two new workloads covering recommendation systems and AI post-training, reflecting the industry’s growing obsession with making increasingly complicated models perform useful tasks. The latest round included 57 cloud-based systems, while 12 configurations employed more than 1,000 accelerators, showing just how enormous AI computing installations have become.
Nvidia’s results suggest its combination of hardware, networking and CUDA software continues to give competitors a difficult target, despite the billions being spent attempting to loosen its grip. The complete benchmark results are available through MLCommons, where customers can examine individual configurations without relying on manufacturers congratulating themselves.







