ReadingResearchRadarInvestment framework
Sign in / Sign up
← Back to reading
TopicDisconfirmAI EditorialUpdated 2026-10-25 15:42

英伟达Rubin性能争议

4 steps1 pieces4 Sources

【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,Rubin每美元总吞吐量约为GB300的67倍,功耗效率提升最高达7倍,远超黄仁勋在GTC 2026上宣称的3倍。

这一差距引发了关于“性能保留”(Sandbagging)的争议:SemiAnalysis认为英伟达刻意低估以管理预期,而部分网友质疑该机构立场偏颇或翻炒旧闻。核心分歧在于测试条件——SemiAnalysis基于预发布软件及特定高并发场景(170 TPS),而实际部署中多数服务商处于60-100 TPS区间,此时优势收窄至1.4至3倍。此外,Rubin Ultra因封装技术难题被曝缩水,进一步加剧了市场对英伟达执行力的担忧。读者需关注后续量产版本的独立基准测试结果,以验证真实性能边界。

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. 2026-09-14

    SemiAnalysis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 ↗
  4. 2026-06-30

    SemiAnalysis发帖称英伟达4芯片版Rubin Ultra因封装翘曲问题被取消,新版规模减半;引发网络对其是否做空英伟达及消息时效性的争议。

    wallstreetcn.com ↗

What is still unsettled

Sign in / Sign up
ReadingResearchRadarInvestment framework
  • 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?
  • Sources

    • newsletter.semianalysis.comSemiAnalysis: Vera Rubin NVL72 Agentic Inference Benchmark↗
    • developer.nvidia.comNVIDIA Developer Blog: Vera Rubin Performance per Watt↗
    • siliconangle.comSiliconANGLE: Nvidia debuts Rubin chip at CES↗
    • wallstreetcn.com华尔街见闻: SemiAnalysis称Rubin Ultra遭取消引热议↗

    Coverage in this topic

    • Compute Research Firm SemiAnalysis Tests Rubin Agentic Inference: Performance per Dollar is 67x that of GB300, Power Efficiency Exceeds NVIDIA's Official Claims2026-10-25News

      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.