Back to news
marketNVIDIA2026-11-06

Nvidia published Vera CPU agent benchmarks and an HBM memory strategy on November 5

In a 100 megawatt data center scenario Nvidia said Vera CPU supports 1.6 times as many agents per watt versus the fastest x86 CPU, and that Vera Rubin NVL72 delivers up to 30 times more agent throughput per dollar while adding only about 3 percent more HBM capacity per GPU.

On November 5, Nvidia published a blog post with measured data for the Vera CPU on agent workloads. Nvidia said that unlike traditional generative AI, agents must not only generate answers but also search for information, query databases, write and test code and actually execute tasks, and those steps rely heavily on the CPU. According to Nvidia's comparison in a 100 megawatt data center scenario, Vera CPU supports 1.6 times as many agents per watt compared with the fastest x86 CPU, and delivers 2 times faster agent execution compared with the most energy-efficient x86 CPU.

Several early partners published their own results. Perplexity said running agent coding tasks on Vera is 1.5 times faster than on conventional CPUs and plans to put it into production, while DeepInfra said Vera is up to 2.2 times faster than the best x86 processors in agent orchestration tasks. For concurrency, Daytona used Vera to start 1000 agent sandboxes in 10.8 seconds and completed 12.7 tasks per second at 2000 concurrent sandboxes, where the x86 platform required 88.4 seconds.

Nvidia also published a technical explainer the same day noting that as model parameter counts, context lengths and agent applications grow quickly, GPU memory faces unprecedented pressure, with HBM supply tightening and prices rising for the foreseeable future. To address the AI memory bottleneck, Nvidia introduced a full-stack co-design approach treating GPU, CPU, network, storage and software as one system. Citing the SemiAnalysis AgentX benchmark, Nvidia said that at 125 tokens per second per user, Vera Rubin NVL72 delivers up to 30 times more agent throughput per dollar than GB300 NVL72, while adding only about 3 percent more HBM capacity per GPU.

英伟达Vera CPUHBM智能体吞吐量全栈协同设计