Nvidia Vera CPU impresses in early Nvidia-sanctioned benchmarks

Daniel Sims

Posts: 2,468   +74
Staff
Highly anticipated: Primarily known for its GPUs and more recently AI infrastructure, Nvidia is set to enter the mainstream CPU market with its Arm-based Vera platform later this year. The company claims Vera delivers significant performance gains over its earlier Grace processors and x86 chips from AMD and Intel. While early tests from Phoronix show encouraging results, Nvidia is still withholding crucial details.

Michael Larabel from Phoronix recently called Nvidia's Vera datacenter CPU the fastest Arm Linux processor he has tested in the outlet's 22-year history. However, since he conducted the benchmarks in a controlled environment at the company's Santa Clara headquarters, the results can hardly be considered definitive.

Most synthetic tests place Vera well ahead of Nvidia's previous Grace CPU and Intel's Xeon server chips, but the competition with AMD's Epyc processors appears much tighter. While Vera clearly outperforms most of Team Red's lineup, AMD's top Epyc models, the 9755 and 975F, trade blows with it. Vera's biggest advantages appear in certain streaming and 7-Zip compression workloads.

However, the one-day comparison was limited to Intel's Xeon 6980P CPUs and a handful of AMD Epyc models, with no AmpereOne review units available. Crucially, other Arm-based CPUs from Apple and Qualcomm were also absent.

Furthermore, Phoronix was unable to monitor power consumption, one of Arm's primary advantages over x86, because Nvidia is still optimizing Vera's power management. Testing was also limited to Nvidia's intended use cases on pre-production silicon. Comprehensive benchmarks will likely have to wait until the CPUs are ready for launch later this year.

Nvidia unveiled Vera at GTC 2026 in March. Designed primarily for AI and analytics workloads, the 88-core processor represents the company's most significant entry into the CPU market to date.

Team Green has largely focused on dominating the GPU market, the primary driver behind the AI boom. Even its previous datacenter CPU, Grace, was mainly sold together with GPUs. Another major shift with Vera is the move away from Arm-designed Neoverse cores in favor of Nvidia's in-house Olympus cores.

While Vera marks the latest step in Nvidia's rise into the server CPU market, another mysterious product, the Nvidia N1X, is expected to bring the company into the consumer processor space soon, starting with laptops later this year.

Little is known about the Arm SoC, but its integrated GPU performance is said to rival the latest mid-range desktop graphics cards. A full reveal is expected at Computex 2026 next week.

Permalink to story:

 
Honestly those benches seem kind of underwhelming, given the resource Nvidia has to throw at this sort of thing. Only beating the nearly two year old competition by 11% in benchmarks you choose is not the best look, especially since AMD has Zen 6 server chips ramping up and likely to be a available by the time Nvidia has these Vera ones ready.
 
Only slightly beating an Epyc 9575F, despite Nvidia controlling what benchmarks were run, it makes no sense why Phoronix agreed to test it. Also no mention of power consumption, but that probably doesn't matter when datacenters get unlimited funds to buy and run as many servers as they want with the promise that AI will do something useful someday.
 
No they don`t. AMD stock rose by 7% today and Nvidia went down about 3% just because of this news.

Vera vs 9575F

88/64 == 1.375 more cores

26.40/ 23.25 == 1.135 more performance

37.5% more cores... for 13.5% more performances over a 2 years old gen...

I'm impressed... ROFL...
 
No they don`t. AMD stock rose by 7% today and Nvidia went down about 3% just because of this news.

Vera vs 9575F

88/64 == 1.375 more cores

26.40/ 23.25 == 1.135 more performance

37.5% more cores... for 13.5% more performances over a 2 years old gen...

I'm impressed... ROFL...
To be fair, they also put 2x9575F… which is 128 cores… which is why there’s something fishy since the performance difference was almost nonexistent…
 
Back