ReadingResearchRadarInvestment framework
Sign in / Sign up
← Back to reading
NewsTopic · 英伟达Rubin性能争议Structural2 min readDisconfirmAI Editorial

Compute Research Firm SemiAnalysis Tests Rubin Agentic Inference: Performance per Dollar is 67x that of GB300, Power Efficiency Exceeds NVIDIA's Official Claims

Compute research firm SemiAnalysis released the first benchmark results for the NVIDIA Vera Rubin NVL72 platform under agentic inference workloads on September 14, 2026. At a specific workload of 170 TPS, Rubin's total throughput per dollar was approximately 67 times that of GB300; in the 60 to 100 TPS range actually deployed by most service providers, this advantage narrowed to 1.4 to 3 times. Regarding power efficiency, SemiAnalysis measured a peak improvement of 7 times token throughput per megawatt, whereas NVIDIA CEO Jensen Huang claimed only 3 times at GTC 2026. SemiAnalysis estimated that Rubin generates approximately 42% more model profit per gigawatt-year than Blackwell Ultra. All data are based on pre-release software.

2026-09-15 06:08 GMT+8

Compute research firm SemiAnalysis released the first benchmark results for the NVIDIA Vera Rubin NVL72 platform under agentic inference workloads on September 14, 2026, using its proprietary AgentX benchmark to replay real-world agentic traffic across a cluster of thousands of chips. Key findings: At a specific interaction speed of 170 TPS (tokens per second), Rubin using NVFP4 dense inference delivered approximately 67 times the total throughput per dollar compared to GB300 Dynamo TRTLLM, assuming an owned-cost model typical of hyperscale cloud providers. However, SemiAnalysis noted that in the 60 to 100 TPS range where most inference service providers actually deploy, Rubin's performance-per-dollar advantage over GB300 narrows to 1.4 to 3 times.

Power efficiency was another significant finding. NVIDIA CEO Jensen Huang presented data at the 2026 GTC conference stating that Rubin achieves a 3x performance increase per megawatt on trillion-parameter models at around 200 TPS; SemiAnalysis's measurements of pre-release software showed a peak improvement of up to 7x token throughput per megawatt. SemiAnalysis cited previous precedents: at GTC 2024, Huang claimed GB200 NVL72 would be 30 times faster than Hopper, but actual testing results showed it was 98 times faster.

Sign in / Sign up
ReadingResearchRadarInvestment framework

Regarding economic implications, based on its AI Cloud TCO model and data center industry model, SemiAnalysis estimated that at an interaction speed of 75 TPS, with 60% utilization and no model license fees, Rubin generates $159.5 billion in annual revenue and $149.9 billion in model profit per full-scope gigawatt of electricity, while the strongest GB300 configuration (Dynamo SGLang) generates $114.9 billion and $105.3 billion respectively. Thus, Rubin gains approximately 42% more model profit from the same power allocation, with an absolute difference of about $44.6 billion per gigawatt. SemiAnalysis also pointed out that Rubin could reduce token pricing by approximately 28% while still maintaining the same per-gigawatt revenue as GB300, providing operators with options between retaining profits and competing on price.

Key limitations must be clarified: all data are based on pre-release TRTLLM software, and SemiAnalysis itself acknowledges that performance will improve as the software stack matures; the 67x figure appears in extreme conditions near the end of the GB300 TRTLLM measurement curve, and engine labels (TRTLLM vs. SGLang) have a huge impact on multipliers at high interaction speeds—at the same 170 TPS condition, Rubin's multiplier against GB300 SGLang is only 5.56x rather than 62.9x. Additionally, SemiAnalysis thanked NVIDIA team members for assisting in verifying benchmark results, which constitutes vendor-collaborative testing rather than completely independent third-party verification.

Sources

  • newsletter.semianalysis.com↗
Topic

英伟达Rubin性能争议

【背景】2026年1月，美国英伟达公司在CES上发布新一代AI平台Vera Rubin，CEO黄仁勋宣称其推理性能为Blackwell的5倍。然而，算力研究机构SemiAnalysis于9月发布的实测报告指出，在代理推理工作负载下，Rubi…

4 steps1 piecesUpdated 2026-10-25 15:42

How it got here

  1. 2026-01-05英伟达在CES 2026发布Vera Rubin平台，黄仁勋称其NVFP4推理算力是Blackwell的5倍，预计2026年下半年出货。siliconangle.com ↗
  2. 2026-08-24英伟达官方博客引用SemiAnalysis AgentX基准，称Rubin NVL72在DeepSeek V4-Pro负载下每兆瓦吞吐量比GB300高30倍。developer.nvidia.com ↗
  3. 2026-09-14SemiAnalysis发布详细报告，实测显示Rubin在170 TPS工况下每美元性能达GB300的67倍，功耗效率提升7倍，批评黄仁勋在GTC 2026仅宣称3倍是“保留性能”。newsletter.semianalysis.com ↗

What is still unsettled

  • 量产版Rubin在主流服务商实际部署的60-100 TPS区间内，能否维持1.4至3倍的显著成本优势？
  • 英伟达是否会针对SemiAnalysis的高倍率实测数据做出正式回应或调整官方宣传口径？
  • Rubin Ultra缩减后的最终规格是否会影响其在高端训练市场的竞争力？
Read the full topic →
Continue researching →