The MLPerf Inference v6.1 benchmark suite, managed by MLCommons, has released its latest performance results. The update focuses on AI inference throughput, scaling efficiency, and hardware configuration testing.
AMD set a new record using a cluster of 512 Instinct MI355X GPUs. In the DeepSeek R1 test, the system achieved an offline throughput of 2,901,950 tokens/s and a server-side performance of 2,405,310 tokens/s.
NVIDIA submitted results for its GB300 and GB200 platforms using 288-GPU configurations. The GB300 recorded 2,705,130 tokens/s in offline mode and 2,028,030 tokens/s in server mode.
The benchmark also featured the debut of NVIDIA's Vera Rubin VR200 architecture. Testing included 36-GPU and 72-GPU configurations, with the 72-GPU VR200 setup showing a 95% performance increase over the GB300.
