Vera Performance Revealed: Can NVIDIA’s New CPU Beat Intel and AMD?
May 27, 2026
The first independent benchmarks for NVIDIA’s Vera processor have arrived, and it looks like bad news for Intel and AMD. Unlike the Grace CPU, which utilized mostly off-the-shelf Arm cores, Vera features NVIDIA’s own custom Olympus core architecture built on the ARMv9.2 instruction set. Packing 88 cores with support for 176 threads, the platform is tailored specifically to anchor the company’s next-gen Vera Rubin AI platform.
According to testing published at Phoronix, performance results for NVIDIA’s Vera CPU were striking: across a geometric mean of enterprise workloads, Vera outpaced AMD’s top-tier high-frequency EPYC 9575F server chip by roughly 11%. Against Intel’s Xeon 6980P Granite Rapids flagship, Vera delivered a staggering 55% lead, even matching or outperforming dual-socket x86 configurations.
Vera’s secret weapon is its custom monolithic design. Relying on a second-gen Scalable Coherency Fabric that offers 3.4TB/s of bisection bandwidth across the die (married to a DDR5X memory subsystem utilizing SOCAMM2 modules), Vera sidesteps the latency penalties that affect AMD and Intel’s multi-chiplet designs.
The setup yields 1.2TB/s of peak memory bandwidth, allowing Vera to maintain a 90% sustained memory bandwidth utilization under the grueling STREAM TRIAD test—the highest percentage ever tested by Phoronix. The architecture also compiled a default Linux kernel in a record-breaking 20 seconds, proving its muscle in standard computing tasks like compilation, file compression, and database management. Generationally, it marks a 1.6x leap over Grace.
However, these results should be taken cautiously. Even though the benchmarking was handled independently, it was restricted to an NVIDIA-approved sandbox of workloads. Testing was limited to specific data center-centric tasks like Python, Java, and database execution where Vera’s immense memory bandwidth could shine, so it remains to be seen how the custom Arm cores handle other, highly unoptimized legacy server tasks or complex scientific computing where x86 traditionally dominates. Furthermore, Vera’s 450-watt thermal design envelope, coupled with a 50-watt draw for its 768GB memory pool, shows NVIDIA isn’t afraid to push power consumption.
Nonetheless, Wall Street and industry analysts are already projecting that NVIDIA’s Vera and Grace processors are going to be hot, fetching $20 billion on their own, and capturing a sizable chunk of the estimated $200 billion data center addressable market.
