NVIDIA Vera CPU Benchmarks: Olympus Cores Delivering The Best Performance Ever Seen On ARM
Across the variety of different workloads that were permitted for testing on this initial round of NVIDIA Vera benchmarking, Vera exceeded my expectations in never seeing an ARM64 processor compete so well against the x86_64 competition. On a geo mean basis, the NVIDIA Vera delivered 10% better performance than the AMD EPYC 9575F 5.0GHz high frequency processor. For gen-on-gen compared to Grace, Vera was coming in at 1.63x the performance geo mean. Over a single Intel Xeon 6980P as Intel's current flagship Granite Rapids processor, NVIDIA Vera delivered 1.55x the performance.
Going into this, I didn't really know what to expect of NVIDIA's Vera with the new Olympus cores. But in the end I was left realizing this is the most formidable competition to Intel and AMD x86_64 processors ever realized. NVIDIA Vera is much more performant than what we have seen out of the likes of Ampere Computing or the custom in-house ARM solutions at public cloud providers like Google Compute Engine and Microsoft Azure. NVIDIA Vera is actually competitive to the latest AMD EPYC and Intel Xeon processors across a swath of workloads. Granted, caveats applied, with NVIDIA limiting the scope of the initial benchmarks to their intended markets and target use-cases, but at least it was diverse enough for some broad realizations and these workloads for the most part are ones I have been running for years in other benchmark pieces. Today's results is just a small sub-set of my typical preferred approach to benchmarking of saturation bombing given today's diverse workloads and a wide swath of user/reader interests, but hopefully we'll be able to do that for fun as Vera ramps up later in the year.
There does also remain open questions around power efficiency / performance-per-Watt with not being permitted to record the CPU power consumption numbers. Though at this early, pre-production stage they may not be too relevant to how customers will find the power tuning with production servers later in the year. So that we will hopefully have a better idea about in the months ahead. Similarly, for those curious about Vera's performance in agentic AI benchmarks, stay tuned for some interesting news this summer.
NVIDIA Vera should be quite interesting when it comes to power usage. The Vera CPU had a 450 Watt TDP while the LPDDR5X memory is only around 50 Watts all while delivering very impressive memory bandwidth. So 500 Watt while the top AMD EPYC Turin and Xeon Granite Rapids CPUs had 500 Watt TDP for the CPU alone. DDR5 ECC RDIMM memory or MRDIMMs with Xeon Granite Rapids can consume much more than 50 Watts when loading all 12 memory channels... For some Xeon 6 MRDIMM/DDR5 memory power (and performance) numbers see this earlier Xeon 6 article. So it should be a win for Vera in overall "wall power" numbers with better power efficiency but no data to share today. And as we saw from the Stream memory benchmark numbers, the LPDDR5X memory was even outpacing the MRDIMMs with Granite Rapids.
So while I don't want to draw any definitive product recommendations/conclusions right now given my limited time with Vera and noted restrictions on selection of benchmarks run, what I can say is that based on the available testing this is definitely the most performant ARM Linux server processor I have ever tested. And Phoronix turns 22 years old in less than two weeks. I have been performance benchmarking on ARM Linux back to the Calxeda server days and other interesting adaptations over the years like solar-powered ARM clusters in trash cans. For the workloads tested, the NVIDIA Vera CPU with Olympus cores proved competitiveness inline with current-generation AMD EPYC 9005 "Turin" processors. NVIDIA Vera more comfortably was beating the top-end Intel Xeon 6980P "Granite Rapids" processors nearly across the board, even with the Xeon 6980P being paired with MRDIMM-8800 memory.
The timing of NVIDIA Vera's ramp and pricing will also play an important role. Vera is ramping later this year. AMD EPYC Venice based on Zen 6 is also expected to launch this year while the exact timing and production ramp are not yet publicly known. EPYC Venice is expected to deliver some significant gains over EPYC Turin, so they may command the lead once they have ramped up. Meanwhile Intel Xeon Diamond Rapids will probably be out sometime in 2027. With AMD EPYC Venice it will probably capture the lead for diverse workloads outside of the designed workloads of Vera, but what perhaps is going to be more significant is how quickly NVIDIA is able to introduce Vera's successor. If NVIDIA continues with their aggressive hardware release cycles and can innovate quicker than AMD's roughly two year cadence between EPYC launches, they could provide some significant, well-timed performance leadership if they are able to build quickly off this excellent foundation of Olympus.
At this stage the pricing also remains a big unknown and how much near-term availability of NVIDIA Vera outside of the hyperscalers, AI companies, and other dominant players. Vera's ramp will be very interesting to see it play out as well as server pricing over the months ahead.
Given the strong per-core performance out of Olympus, desktop/workstation users may be eager over the dreams of a NVIDIA desktop/workstation class processor. Unfortunately, that sounds unlikely. Due to various TCO factors, it doesn't sound like NVIDIA is pursuing any lower core count Vera processors, but we can always hope for something further out in the future.
Beyond the aggressive performance out of NVIDIA Vera with its competitiveness to x86_64 and excellent gen-on-gen performance compared to Grace, the other enlightening experience from my time at NVIDIA was seeing the great upstream open-source support for Vera. Given they aren't relying on ugly Device Tree files and complying nicely with the various Arm standards, the upstream kernel support for NVIDIA Vera is in good shape and in turn for modern, prominent AArch64 Linux distributions. This was a gray box going into it with not really knowing what to expect or where any standards would be broken or bespoke driver solutions required, but with the Olympus cores it's great to see everything able to run nicely on a mainline Linux kernel. Additionally, NVIDIA having upstreamed their Olympus core support to the GCC and LLVM Clang compilers well in advance of launch.
Thanks again to NVIDIA inviting me out to conduct these early, exclusive Vera performance benchmarks. Lastly, while at NVIDIA and meeting with various executives and with the tease, you can be sure I prompted them on various open-source topics outside the realm of Vera, including matters like the Nova upstream open-source graphics driver that continues seeing significant NVIDIA engineering contributions. Stay tuned for separate follow-up articles on those as well as a few other Vera benchmark articles still to come.
If you enjoyed this article consider joining Phoronix Premium to view this site ad-free, multi-page articles on a single page, and other benefits. PayPal or Stripe tips are also graciously accepted. Thanks for your support.