Nvidia’s 88-core Vera server CPU often matched or beat comparable Intel and AMD processors in Tom’s Hardware’s tests, with particularly strong results in Python and memory bandwidth. The catch is that these are workload-focused clues, not a standardised test of a complete AI agent loop.
NVIDIA Watch analysis
What happened
Tom’s Hardware tested a dual-socket Vera system against Intel Xeon 6980P and AMD EPYC 9965 and 9755 processors. The report says Vera performed strongly in Python and delivered up to 1.2 TB/s of aggregate memory bandwidth. It also often matched or beat the rival chips in general-purpose server workloads, while x86 processors could lead in some workloads that benefit from their core density.
The testing included established workloads such as Python, code compilation and database performance, viewed through the needs of AI agents. The report says standardised benchmarks for complete agentic workflows are not yet available. Its SPEC CPU 2026 results are estimates, not results submitted to SPEC.
Why it matters
AI agents do more than run models on GPUs. CPUs also manage agent state, coordinate work and run tools such as code and database queries. Vera’s results suggest Nvidia may have a credible server CPU for some of that supporting work, not just an accelerator to sit beside someone else’s processor.
But a strong result in a selection of relevant workloads is not the same as proving an advantage across real agent systems. Buyers still need to weigh the workload, platform and operating costs, while the benchmark field for end-to-end agent work catches up.
Our read
This is a promising first showing for Nvidia’s custom CPU, with a useful reminder that AI infrastructure is more than GPUs in an expensive waistcoat. The evidence points to competitive performance, not a universal win, and the missing full-loop benchmark matters. Server buyers should treat Vera as a serious option to compare against their own workloads, rather than a reason to retire the spreadsheet.
What to watch
- Whether independent tests reproduce Vera’s strong Python and memory-bandwidth results.
- How Vera performs in standardised, end-to-end agentic workloads when those benchmarks arrive.
- How its performance and platform costs compare for specific server deployments.
Discussion spark: Should server buyers take Vera’s strong results on selected AI-adjacent workloads as enough to consider switching, or wait for standardised benchmarks of complete agent workflows?
Sources and evidence
- Benchmarking Nvidia’s Vera CPU for the data center – standalone Vera tested against Granite Rapids-AP and EPYC Turin (30 October 2026, 11:00 UTC)
Independent WittyWires tracker for public updates about NVIDIA. Not affiliated with or endorsed by NVIDIA; this is not an official account.