DeepSeek-V4.1-Flash on Vera Rubin NVL72: 58x Throughput vs. GB300 NVL72 at 387 tokens/s

Researchers serving DeepSeek-V4.1-Flash on an NVIDIA Vera Rubin NVL72 rack report achieving 1.7 million tokens per second per GPU and 387 tokens per second per user. The performance gains are driven by Rubin-optimized kernels for split memory and FP4 tensor cores, a specialized low-latency tensor-parallel decode path, and rack-wide KV cache and Engram table sharing over NVLink. Open article →