Key points
- Tom's Hardware tested the standalone Nvidia Vera CPU against Intel Granite Rapids-AP and AMD EPYC Turin processors.
- The Vera chip delivered top-tier Python performance and high memory bandwidth, critical for agentic AI workflows.
- Nvidia's custom Olympus core shows significant improvements over the previous Grace CPU architecture.
Tom's Hardware reported that Nvidia’s new Vera data center CPU demonstrates strong performance in Python and memory bandwidth during independent benchmarking against Intel’s Granite Rapids-AP and AMD’s EPYC Turin processors.reported The outlet tested the standalone Vera chip to evaluate its capabilities in current AI infrastructure workloads.Independent tests show Nvidia's new chip is very fast at running Python code and moving data, which helps AI programs work better.Tom's Hardware benchmarks indicate the Vera CPU's custom Olympus core provides superior memory bandwidth and Python throughput compared to x86 alternatives.
Evidence: reported
Benchmarking Nvidia’s Vera CPU for the data center — standalone Vera tested against Granite Rapids-AP and EPYC Turinaifleet-tomshardware
How we got here
Nvidia has spent years building AI infrastructure, positioning the Vera CPU as a response to rising demand for data center processing power. The company markets the chip specifically for the shift toward agentic AI workflows, although Tom's Hardware noted that standardized end-to-end agentic benchmarks for CPUs do not yet exist.
What is new
The Vera CPU uses Nvidia’s first custom core, named Olympus, which is built on the Armv9.2-A architecture.reported This design includes two threads per core through spatial multithreading. Tom's Hardware described the Olympus core as a stark improvement over the Grace CPU, which relied on an off-the-shelf Arm core design.
Evidence: reported
Benchmarking Nvidia’s Vera CPU for the data center — standalone Vera tested against Granite Rapids-AP and EPYC Turinaifleet-tomshardware
Reading between the lines
The 88-core Vera processor offers up to 1.2 TB/s of aggregate memory bandwidth and 14 GB/s per core.reported It supports 88 PCIe 6.4 lanes when used alone, or 96 lanes in Vera Rubin configurations, with bifurcation down to x2 and CXL 3.1 support.reported These specifications aim to address the high-speed data transfer needs of complex AI agents.
Evidence: reported
Benchmarking Nvidia’s Vera CPU for the data center — standalone Vera tested against Granite Rapids-AP and EPYC Turinaifleet-tomshardware
Evidence: reported
Benchmarking Nvidia’s Vera CPU for the data center — standalone Vera tested against Granite Rapids-AP and EPYC Turinaifleet-tomshardware
The open questions
Tom's Hardware found that while Vera excels in backend, data-rich operations, x86 CPUs often win on core density for frontend workloads.reported The outlet stated that Vera’s performance in agentic-adjacent tasks is not surprising, but its competent general-purpose server performance is. The benchmarks suggest Vera trades core density for strong memory and Python execution, raising questions about optimal workload placement.
Evidence: reported
Benchmarking Nvidia’s Vera CPU for the data center — standalone Vera tested against Granite Rapids-AP and EPYC Turinaifleet-tomshardware
Why it matters
- Model evaluators can use the Python performance data to estimate inference speeds for Python-heavy AI agents.
- The memory bandwidth figures help assess whether Vera can handle large database queries common in agentic workflows.
- Comparing Vera to x86 rivals provides context for cost-performance tradeoffs in mixed server environments.
- The lack of standardized agentic benchmarks means current results are indicators, not definitive proofs of superiority.
Topics: Nvidia, CPUs, benchmarks, Arm architecture, AI infrastructure
Updated
Then / now / next
Nvidia’s Grace CPU used an off-the-shelf Arm core design.
The new Vera CPU uses a custom Olympus core with improved performance metrics.
Evaluators should watch for the development of standardized agentic AI benchmarks.
Story timeline
- Then Nvidia’s Grace CPU used an off-the-shelf Arm core design.
- 30 Oct 2026 Benchmarking Nvidia’s Vera CPU for the data center — standalone Vera tested against Granite Rapids-AP and EPYC Turin
- Now The new Vera CPU uses a custom Olympus core with improved performance metrics.
- Next Evaluators should watch for the development of standardized agentic AI benchmarks.
The strongest doubt
Counterpoint from the sources
Tom's Hardware noted that the idea of developing a CPU for a demand shift that arose only six months ago is reaching, suggesting the marketing narrative may be overstated.
Ask the article
Local keyword search of this article; no generated answers.Did this article give you the evidence you needed?
Your choice stays on this device. No public tally.Terms in this story
Olympus
Nvidia's first custom CPU core design used in the Vera processor.
Spatial multithreading
A technique allowing two threads per core in the Vera CPU architecture.
Agentic AI
AI workflows where agents perform complex tasks, often requiring high memory bandwidth.
Questions readers ask
How does Nvidia Vera CPU compare to Intel and AMD chips?
Tom's Hardware reported that Vera matches or beats comparable AMD Turin and Intel Granite Rapids-AP chips in general server workloads, while leading in Python and memory bandwidth.
What architecture does the Vera CPU use?
The Vera CPU uses Nvidia’s custom Olympus core based on the Armv9.2-A architecture with spatial multithreading.
Are there standardized benchmarks for agentic AI CPUs?
No, Tom's Hardware stated that standardized end-to-end agentic benchmarks focused on the CPU do not currently exist.
Connected stories
Reading receipt
- Tom's Hardware tested the standalone Nvidia Vera CPU against Intel Granite Rapids-AP and AMD EPYC Turin processors.
- The Vera chip delivered top-tier Python performance and high memory bandwidth, critical for agentic AI workflows.
- Nvidia's custom Olympus core shows significant improvements over the previous Grace CPU architecture.
Still open: How will Vera perform in standardized agentic benchmarks once they are developed?
Still open: Will the tradeoff between core density and memory bandwidth favor Vera in mixed workloads?
