Nvidia Benchmarks Vera CPU: 60% More Agents per Watt Than Fastest x86
Nvidia published benchmark results for its Vera CPU on agentic workloads, claiming that in a 100 MW data center scenario it supports 1.6x more agents per watt than the fastest x86 CPU and delivers 2x agent speed versus the most efficient x86 part. Early partners including Perplexity, DeepInfra, ClickHouse, Daytona and Modal reported speedups in coding, orchestration and concurrent sandbox scenarios, reframing the CPU side as a scaling bottleneck for agent deployments.
Full Report
IT Home November 6 news, NVIDIA yesterday (November 5) published a blog post, revealing measured data for the Vera CPU (to be released on March 16, 2026) on agentic workloads, claiming that it can simultaneously achieve industry-leading agent execution speed and throughput per watt in a single processor, helping data centers run more agents under limited power conditions.
NVIDIA stated that, unlike traditional generative AI, agents not only generate answers but also search for information, query databases, write and test code, and actually execute tasks. These steps rely heavily on the CPU and require repeated switching between the CPU (responsible for execution) and the GPU (responsible for inference). NVIDIA believes that as enterprises deploy agents at scale, the core bottleneck of data centers is shifting from raw compute power to a balance of "speed, concurrency capability, and energy consumption."
According to NVIDIA's comparative testing in a 100-megawatt data center scenario, the Vera CPU can support 1.6 times the number of agents per watt compared with the fastest x86 CPU; compared with the most energy-efficient x86 CPU, agent speed is doubled while maintaining leading energy efficiency.
Citing the blog post, IT Home notes that several early partners have published test results:
- Perplexity says that using Vera to run agentic coding tasks is 1.5 times faster than traditional CPUs, and it plans to put it into production systems;
- DeepInfra says that in agent orchestration tasks, Vera is up to 2.2 times faster than the best x86 processors;
- ClickHouse says that its standard software can scale nearly linearly across 176 cores without specialized optimization, and it has entered the ClickBench hardware rankings.
- In concurrency scenarios, Daytona used Vera to launch 1,000 agent sandboxes in 10.8 seconds and achieved 12.7 completed tasks per second under 2,000 concurrent sandboxes, whereas the x86 platform required 88.4 seconds.
- Modal reported that Vera's per-core sandbox throughput reached up to 2 times that of comparable x86 hardware.