The development

What happened

NVIDIA reported Vera Rubin NVL72 results for AI training and agentic inference on 16 November. The company claims up to twice the training performance per GPU versus GB300 NVL72, but its highlighted inference-efficiency comparison uses a provisional Vera Rubin result that still awaits MLCommons verification.

Key facts

  • MLPerf Training measures time to reach a target accuracy. NVIDIA reports results across all benchmarks in the round and up to 2x training performance per GPU over GB300 NVL72; these are company-reported benchmark results, not a universal speedup.

  • For DeepSeek-V3 training, NVIDIA reports about 97% scaling efficiency from 256 to 1,024 Vera Rubin GPUs, with time falling from 16 minutes to about four. That example does not establish the same scaling behavior for every model or cluster.

  • Its Kimi K3 inference example claims more than 10x throughput per watt at a shared minimum of 50 tokens per second per user. NVIDIA labels the Vera Rubin result provisional and pending MLCommons verification; the GB300 result is described as verified.

  • The comparison uses 32 Vera Rubin accelerators with 1,024 concurrent queries versus 48 GB300 accelerators with 48 queries, at reported operating power of 100.89 and 60 kW. These are different system configurations, not an identical-card test.

  • The inference footnote dates its results and retrieval to 9 November, while the training-results footnote is dated 16 November. TideReach has not independently re-run the tests or checked the underlying leaderboard submissions.

The context

Why it matters

TideReach assessment: training time, total serving throughput and response speed answer different questions. The efficiency claim is about system throughput per watt at a stated response-speed floor, not each request becoming ten times faster or an agent becoming more capable. Buyers still need workload, latency, quality and power comparisons that fit their deployment; provisional results can change after review.