Phoronix benchmarks NVIDIA Vera CPU against Intel and AMD x86 flagships

What's new

On May 26, 2026, Phoronix published the first independent benchmark suite for the NVIDIA Vera CPU, the data center Arm part that anchors the Vera Rubin platform. Reviewer Michael Larabel tested a single-socket Vera with 88 custom NVIDIA Olympus cores on the Armv9.2 ISA, rated at a 450-watt thermal design power with under 30 watts of memory power, paired with a second-generation LPDDR5X subsystem delivering up to 1.2 TB/s of memory bandwidth. Across code compilation, file compression, video transcoding, Python, Java, and database workloads, Vera posted a 1.6x geometric-mean gain over the prior-generation NVIDIA Grace CPU and a 1.5x overall lead versus a current 128-core x86 processor. On Linux kernel compilation, single-socket Vera finished in 20 seconds, the fastest result Phoronix has recorded on that test, and roughly 2x faster per core than the 128-core x86 part. STREAM TRIAD sustained 90% of Vera's rated peak bandwidth, the highest peak-bandwidth utilization Phoronix has measured on any CPU, and yielded over 4x the memory bandwidth per core compared with traditional x86. On a geometric-mean basis, Larabel reported Vera 10% ahead of AMD's EPYC 9575F running at 5.0 GHz. NVIDIA covered the results in its own blog the same day. Source: NVIDIA Blog, "NVIDIA Vera CPU Is 'Packing a Heavy-Hitting Punch' Against Competition," and Phoronix, "NVIDIA Vera Benchmarks."

Why it matters

Vera is the first NVIDIA data center CPU built around fully custom NVIDIA Olympus cores rather than off-the-shelf Arm Neoverse IP, and Phoronix's numbers are the first third-party data point on whether the Olympus design competes outside vendor-controlled MLPerf-style submissions. A 1.5x geomean lead over a 128-core x86 part on broad developer workloads, and a 10% edge over EPYC 9575F on a per-socket basis, is a meaningful shift: the host CPU has historically been an x86 default in AI factories regardless of which accelerator sat in the slot. The LPDDR5X bandwidth result reinforces NVIDIA's design pitch, which is that agentic inference is rate-limited by host-side branch-heavy runtimes, tool calls, sandbox execution, and database queries that sit between accelerator runs and depend on memory bandwidth per watt, not raw socket count.

Caveats

The review is a single-vendor disclosure: Phoronix tested hardware coordinated with NVIDIA, and the comparison x86 parts were stock-clock production silicon rather than competitive bring-up tunes. Larabel ran a single-socket Vera; dual-socket scaling, real-world coherency under sustained NUMA load, and behavior in shared multi-tenant AI factory configurations are not in scope. Workload coverage skews to compile, compression, and database benchmarks rather than the full agentic-inference pipelines NVIDIA cites in its marketing, and the EPYC 9575F comparison is one SKU rather than a swept curve across Turin or Granite Rapids parts. Vera is not yet generally available; NVIDIA says partner systems in dual- and single-socket, air-cooled and liquid-cooled configurations are scheduled for the second half of 2026, and pricing has not been disclosed. Source: NVIDIA Blog, May 26, 2026.