AI agents rely on CPUs for much more than generating answers. They search for information, query databases, write and test code, and perform actions while moving between GPU inference and CPU execution. NVIDIA says its Vera CPU is designed to help AI infrastructure providers run more agents within tight power limits, delivering high agent performance and capacity per watt in a single processor.
Why CPUs matter for AI agents
As enterprises deploy more AI agents, data centers need to complete each agent’s work quickly while supporting many agents simultaneously. Because AI factories have limited power, infrastructure operators must maximize the amount of work performed per watt.
NVIDIA says Vera is intended to address both requirements. The company describes it as a CPU capable of delivering fast agent performance and high capacity per watt, allowing providers to support a range of agent workloads without combining faster and denser x86 processors to balance speed and capacity.
NVIDIA’s performance and efficiency claims
In a comparison involving 100-megawatt data centers, NVIDIA said a single Vera CPU outperformed the fastest and most energy-efficient x86 CPUs tested.
- Compared with the fastest x86 CPU, Vera delivered industry-leading agent speed and supported 1.6 times more agents per watt.
- Compared with the most energy-efficient x86 CPUs, agents were up to twice as fast, according to NVIDIA.
These results are NVIDIA’s claims based on its comparisons. The company says early partners have reported similarly strong performance in operational-style tests.
Early partner results
- Perplexity said Vera enabled agent coding tasks to run 1.5 times faster than on traditional CPUs. The company plans to use Vera in its next operational system.
- DeepInfra reported that Vera delivered agent orchestration for production workloads up to 2.2 times faster than the best x86 processors, making it the fastest of the four architectures tested.
- ClickHouse said its standard binary, without Vera-specific tuning, scaled almost linearly across 176 cores and topped the ClickBench hardware leaderboard.
- DuckDB achieved approximately 1.5 times the median performance of the Intel Xeon 6 baseline in TPC-H SF1000 testing.
- Daytona brought 1,000 agent sandboxes online in 10.8 seconds and achieved 12.7 completed jobs per second with 2,000 concurrent sandboxes. The source comparison lists 88.4 seconds on x86 with Vera.
- Modal reported up to twice the sandbox throughput per core compared with comparable x86 hardware, with greater improvements under heavy loads.
More information is available on the NVIDIA Vera CPU.
Source: blogs.nvidia.com