NVIDIA's Vera Rubin and GB300 DeepSeek V4.1 Flash benchmark results on Artificial Analysis

AINVIDIA has submitted updated Vera Rubin NVL72 configurations for DeepSeek V4.1 Flash to Artificial Analysis. Per GPU, Vera Rubin delivers 5.0x GB300's throughput at 260 tok/s per agent and 1.7x at 80 tok/s per agent. With one agent, it reaches 1,249 tok/s versus 848 tok/s on GB300, a 1.47x gain.

AIArtificial Analysis says DeepSeek V4.1 Flash speed on NVIDIA GB300 NVL72 rose up to 1.5x tok/s per agent over the past month on its AA-AgentPerf benchmark. The gains come from a newer vLLM version, new vLLM kernels for DeepSeek V4.1 on GB300, FlashInfer autotuning, fuller CUDA graph coverage, tuned speculative decoding, and running prefill and decode on the same GPUs in NVIDIA's updated Dynamo + vLLM setup. A single agent now reaches 848 tok/s, up from 596 (1.42x), and throughput per GPU is about 1.6x at roughly 80 tok/s per agent.