MLCommons has released the MLPerf Inference v6.1 benchmark suite, featuring new submissions from AMD, NVIDIA, and Intel. AMD showcased a 512-GPU Instinct MI355X cluster that led performance for DeepSeek R1, while Intel provided data for its Arc Pro B70 GPU.
NVIDIA tested its GB300 and GB200 hardware, with the Vera Rubin VR200 showing significant gains. A 72-GPU VR200 setup proved 95% faster than the GB300 in DeepSeek R1 previews. Meanwhile, Intel Xeon remained the sole standalone CPU entry in the latest suite.
The results highlight intense competition in the AI sector. NVIDIA Blackwell led Llama2 70B tests, though AMD's Instinct platform outperformed the GB200 there. NVIDIA also reported 1.6x gains from software updates, while AMD's MI350P beat the older MI300X.