Tom's Hardware published benchmarks of Nvidia's Vera data center CPU, finding it matched or beat comparable AMD EPYC and Intel Xeon processors in some tests relevant to agentic workflows, though the reviewers note server performance depends on use cases and software optimisation.
The publication tested Vera against AMD's EPYC 9965 and 9755 Turin chips and Intel's Xeon 6980P Granite Rapids-AP on a range of workloads. Nvidia's server CPU, built around its own 'Olympus' core, features 88 cores on the Armv9.2-A architecture and uses replaceable SOCAMM2 LPDDR5X memory to achieve up to 1.2 TB/s of memory bandwidth.
Nvidia markets Vera as an agentic CPU, amid the rise in prominence of agentic CPUs. Standardised benchmarks for the end-to-end agentic loop do not yet exist. Tom's Hardware evaluated Vera on component tasks relevant to these workflows, noting the industry is still defining how to benchmark this work. Google Cloud's VP/GM of AI and computing infrastructure Mark Lohmeyer told the publication the industry needs new benchmarks for agentic workloads similar to those developed for GPU and TPU inference.
The chip delivered its strongest results in Python performance, topping the PyBench and PyPerformance tests, given Python's role in foundational AI agent frameworks. It also demonstrated leading memory bandwidth in the STREAM benchmark and showed strong results in the SPEC CPU 2026 integer throughput suite, where it beat the EPYC 9755 by an estimated 1.5%.
Vera showed weaker cryptography performance and trailed in Zstd decompression, though it offered impressive Zstd compression performance. In frontend tasks, the higher core density of x86 chips was often an advantage. Tom's Hardware noted that Nvidia did not limit the tests it could run or place any stipulations on the coverage.
Analysis
Why This Matters
- An Nvidia CPU directly competing with AMD and Intel's established server chips represents a major shift in the data center landscape, driven by the AI infrastructure buildout.
- The results validate that Arm-based chips can match or beat x86 in critical AI and general-purpose data center tasks, potentially breaking the x86 duopoly.
- As agentic AI increases demand for CPU orchestration, the choice of server silicon becomes a strategic decision for enterprises and hyperscalers.
Background
Nvidia entered the server CPU market with Grace, an earlier Arm-based design. Vera represents Nvidia's second generation, featuring its own custom-designed 'Olympus' core rather than an off-the-shelf design. The chip enters a market dominated by Intel's Xeon line and AMD's EPYC line, both using the x86 architecture. Nvidia explicitly frames Vera as a response to the shift toward agentic AI, where CPUs manage the orchestration, tool execution, and multi-agent state.
Key Perspectives
Nvidia: Markets Vera as a CPU specifically designed for the shift to agentic AI, emphasising its memory bandwidth and high concurrency.
Tom's Hardware: Found Vera has strong performance indicators for agentic workflows, especially in Python and memory bandwidth, and matches x86 in general-purpose tasks, but lags in cryptography and compression. They caution that server performance is heavily dependent on the specific use case and software stack.
What to Watch
- Arrival of next-generation x86 competitors, particularly AMD's Venice and Intel's Diamond Rapids, which promise significantly improved memory bandwidth.
- Development of standardised industry benchmarks for full agentic AI loops, which will provide a clearer comparison for the class of CPU.
- Adoption decisions from major hyperscalers and cloud providers, who are the primary customers for chips of this class.
Checked
12 claims in this story checked against 1 source: 11 supported.
Show each claim
- SupportedNvidia Vera posts strong results in workloads relevant to agentic computing.Latest from Tom's Hardware: “Vera has strong indicators about its usefulness in agentic workflows.”
- SupportedTom’s Hardware found Vera matched or beat comparable AMD and Intel chips in some tests.Latest from Tom's Hardware: “often managing to match or beat the comparable AMD Turin and Intel Granite Rapids-AP chips.”
- SupportedTom’s Hardware said its testing showed commanding memory bandwidth and Python performance.Latest from Tom's Hardware: “It offers chart-topping performance in Python, for instance, which makes up much of the backbone of the AI software ecosystem. It also shows commanding memory bandwidth”
- SupportedTom’s Hardware tested Vera against AMD’s EPYC 9965 and 9755 and Intel’s Xeon 6980P.Latest from Tom's Hardware: “Tested Processors 2P Nvidia MGX Vera reference platform ... Intel Xeon 6980P ... AMD EPYC 9965, EPYC 9755”
- SupportedVera is built around Nvidia’s own Olympus core, has 88 Armv9.2-A cores, and uses replaceable SOCAMM2 LPDDR5X memory with up to 1.2 TB/s bandwidth.Latest from Tom's Hardware: “The 88-core CPU is built on the Armv9.2-A architecture”
- SupportedNvidia markets Vera as an agentic CPU, as agentic CPUs have grown more prominent.Latest from Tom's Hardware: “Given the rise in prominence of agentic CPUs”
- SupportedStandardized benchmarks for the full end-to-end agentic loop are not yet available.Latest from Tom's Hardware: “Standardized benchmarks that look at the full agentic loop aren’t available yet.”
- SupportedTom’s Hardware tested component tasks relevant to agentic workflows while the industry is still developing ways to benchmark the full workload.Latest from Tom's Hardware: “we can still look at CPU performance through the lens of an agentic workflow”
- SupportedGoogle Cloud VP/GM Mark Lohmeyer told Tom’s Hardware that the industry needs agentic-workload benchmarks like those emerging for GPU and TPU inference.Latest from Tom's Hardware: “I think this is happening now for GPUs and TPUs for inference, for example. I think we're going to need the same type of industry-level benchmarks for agentic workloads”
- Partly supportedVera topped the PyBench and PyPerformance tests.Latest from Tom's Hardware: “Outside of a couple of misses in web server-focused tests, Vera tops the charts across the board.”
- SupportedVera led the STREAM benchmark for memory bandwidth and beat the EPYC 9755 by an estimated 1.5% in SPEC CPU 2026 integer throughput.Latest from Tom's Hardware: “with our tests, Vera is ahead by 1.5%.”
- SupportedVera showed weaker cryptography performance, trailed in Zstd decompression, and had impressive Zstd compression performance.Latest from Tom's Hardware: “Vera offered impressive compression performance with Zstd but fell behind on decompression.”
