NVIDIA said in a newsroom post1 on Nov. 5 that its Vera CPU delivers 2x faster agent speed than the most energy-efficient x86 CPU and 1.6x more agents per watt than the fastest x86 CPU.

The figures come from what NVIDIA calls a 100-megawatt data center comparison. The post does not name the x86 processors, the workload or the test setup, so the two ratios cannot be checked against specific chips.

NVIDIA argues that agents move between GPU reasoning and CPU execution throughout a task, and that AI factories are power limited, so operators need to maximize work per watt. It says Vera gives providers one CPU for a range of agent workloads, instead of mixing "fast and dense x86 parts" to balance speed and capacity.

What NVIDIA and its partners reported

DeepInfra, an inference cloud provider, cites an NVIDIA public claim of "80% faster agentic CPU performance", which it treats as a 1.8x bar. The NVIDIA post includes no pricing or availability date for Vera. It cites results from six partners, none independently verified, and gives no test configurations for them.

The Vera chip has 88 NVIDIA-designed Olympus cores that are fully Arm v9.2-compatible, according to ClickHouse3. A two-socket system has 176 cores and 352 hardware threads through NVIDIA's Spatial Multithreading.

NVIDIA's post lists these partner results:

  • Perplexity: agentic coding tasks ran 1.5x faster than on "traditional CPUs", and it plans to deploy Vera in an upcoming production system.
  • DuckLabs: about 1.5x the median TPC-H SF1000 performance of an Intel Xeon 6 baseline.
  • Daytona: 1,000 agent sandboxes online in 10.8 seconds, versus 88.4 seconds with x86.
  • Modal: up to 2x sandbox throughput per core versus comparable x86 hardware.

DeepInfra: 1.2x to 2.2x, depending on the baseline

DeepInfra measured four architectures2: NVIDIA Vera, AMD Zen5 (Turin), Intel Granite Rapids and Intel Sapphire Rapids. It replayed captured traffic from a production agent at fixed pacing, with no live model tokens generated. Each chip ran on 20 physical cores, one thread per core, on a single NUMA node. DeepInfra said it locked the comparison basis, orchestration-step latency on bare metal per core, before measuring Vera.

Vera's latency was 29 ms against 64 ms for Sapphire Rapids, which DeepInfra put at 2.2x. Against Granite Rapids, the best Intel chip, the margin was 1.76x. Against AMD Zen5, the best x86 result, it was 1.2x. DeepInfra called "fastest of all four" the honest framing.

NVIDIA's post describes DeepInfra's result as "up to 2.2x speed of the best x86 processor." By DeepInfra's own figures, the 2.2x is against Sapphire Rapids, not the best x86 chip.

DeepInfra describes its testing as independent of NVIDIA's numbers. It also said it received early access to Vera hardware, runs agents in production and is planning around Vera as Vera Rubin platforms become available. Because of that relationship, its results are treated here as partner-run.

On capacity, DeepInfra set a service-level bar of 99th-percentile response at or below six seconds with zero errors. On identical 20-core partitions running secure microVMs, Vera sustained 256 concurrent agents. Incumbents managed 160, and Granite Rapids reached 192 on its best run. DeepInfra said Vera's latency tail at 192 agents matched what incumbents showed at about 128.

DeepInfra also reported that its own production Xeon, tuned for inference throughput, was slower at agent work than a stock cloud instance of the same Granite Rapids silicon, despite higher instructions per clock (2.63 versus 2.44). It attributed this to a deliberate all-core clock cap of about 2.55 GHz. Across all chips, it said microVM isolation was the largest cost of hosting agents. It put that cost at 1.5x to 2x on typical work and about 5x to 13.5x on process spawning, with Arm architectures paying about half the x86 spawn penalty.

ClickHouse: standard binary on NVIDIA-supplied server

ClickHouse said NVIDIA provided it with an early-access server. Using its standard binary with no Vera-specific changes, it said Vera topped the ClickBench hardware leaderboard3 "by a wide margin."

In its microbenchmarks, CPU-bound queries scaled almost perfectly with core count. Efficiency began to decline at 176 cores, the number of physical cores in the machine. A simple filter query stopped scaling at 64 threads, reaching 813 GB/s at 64 threads and 900 GB/s at 256. ClickHouse said that roughly matches the system's physical memory bandwidth, which it gave as 1.0 TB/s to 1.2 TB/s depending on DRAM capacity.