NVIDIA publishes Vera CPU benchmarks: 1.6x more agents per watt

NVIDIA published comparison data for its Vera CPU on November 5, positioning the chip as the answer to a problem that shows up once a company runs many AI agents at once: the work keeps bouncing between GPU reasoning and CPU execution.

The blog post lays out the mechanics plainly. Agents search for information, query databases, write and test code, and take real actions, and each of those steps lands on a CPU. As more agents run in parallel, infrastructure providers have to add capacity without making every individual agent wait longer to get its work done.

In a 100-megawatt data center comparison, NVIDIA says a single Vera CPU outperformed both the fastest x86 CPU and the most energy-efficient x86 CPU. Against the fastest x86 part, Vera enabled 1.6x more agents per watt; against the most efficient x86 part, it delivered 2x faster agent speed with leading energy efficiency.

The pitch is consolidation. Operators get one CPU for a spread of agent workloads instead of mixing fast and dense x86 parts to balance speed against capacity, NVIDIA argues.

Early adopters report production-style numbers. Perplexity ran agentic coding tasks 1.5x faster than on traditional CPUs and plans to deploy Vera in an upcoming production system. DeepInfra measured up to 2.2x the speed of the best x86 processor on agentic orchestration. ClickHouse said its standard binary, with no Vera-specific tuning, scaled near-linearly across 176 cores and topped the ClickBench hardware leaderboard.

Daytona brought 1,000 agent sandboxes online in 10.8 seconds on Vera versus 88.4 seconds on x86, and reached 12.7 completed jobs per second at 2,000 concurrent sandboxes. Modal reported up to 2x sandbox throughput per core versus comparable x86 hardware, with gains that grow under heavy load. DuckLabs cited roughly 1.5x the median TPC-H SF1000 performance of an Intel Xeon 6 baseline.

The framing matters more than any single multiplier. AI factories are power limited, so operators are being asked to maximise work per watt rather than peak compute. NVIDIA is arguing that the useful metric is work completed per watt, not peak compute, and that agent density belongs in procurement decisions alongside raw throughput.

The caveats are the usual ones for vendor benchmarks. All figures come from NVIDIA or from partners with an interest in the result, and the comparison chips are described generically as the fastest and most efficient x86 rather than by SKU. Perplexity's production deployment and independent retesting will be the real test of where Vera actually sits.