MLPerf v6.1 gives AMD, Intel, and NVIDIA different wins in its AI tests. AMD used 512 Instinct MI355X GPUs. DeepSeek R1 hit 2.90 million tokens per second offline. GPT-OSS-120B hit 5.75 million there.
Vera Rubin VR200 made its first test run for NVIDIA. Its 72-GPU result beat the 72-GPU GB300 mark by as much as 95%. A 36-GPU VR200 box also outdid one built with 72 GB200 chips. At eight GPUs, GB300 beat MI355X in Llama 2 70B. It was 17% faster in Llama 3.1-8B.
Intel sent in Arc Pro B70 and Xeon 6 results. Four B70 cards had 128GB VRAM. Xeon 6 showed five SKUs. Xeon was also the only CPU line used alone in a server.
And code tweaks kept changing the scores. NVIDIA said its work raised speed by as much as 1.6x from v6.0. AMD got more from MI355X with no hardware swap. Intel also raised B70 results on its prior four-card system.
Vera Rubin VR200 made its first test run for NVIDIA. Its 72-GPU result beat the 72-GPU GB300 mark by as much as 95%. A 36-GPU VR200 box also outdid one built with 72 GB200 chips. At eight GPUs, GB300 beat MI355X in Llama 2 70B. It was 17% faster in Llama 3.1-8B.
Intel sent in Arc Pro B70 and Xeon 6 results. Four B70 cards had 128GB VRAM. Xeon 6 showed five SKUs. Xeon was also the only CPU line used alone in a server.
And code tweaks kept changing the scores. NVIDIA said its work raised speed by as much as 1.6x from v6.0. AMD got more from MI355X with no hardware swap. Intel also raised B70 results on its prior four-card system.
