TL;DR
- Benchmark Result: Nivida’s Vera CPU beat comparable AMD EPYC and Intel Xeon parts in partially controlled benchmark results.
- Method Limits: Nvidia restricted workloads and blocked power monitoring, so buyers still lack a full independent validation package.
- Rollout Stakes: Oracle plans to deploy hundreds of thousands of Vera CPUs in 2026 as Rubin systems move toward second-half availability.
Nvidia’s Vera CPU finished ahead of AMD EPYC and Intel Xeon in early benchmark results shared by phoronix. Nvidia controlled the workload list for that session and blocked power and frequency monitoring, which leaves buyers without the broader evidence they would usually want before treating a launch-stage server result as settled.
Vera is Nvidia’s next-generation Arm-based server CPU, designed to sit alongside its AI accelerators in large datacenter systems. Unlike AMD EPYC and Intel Xeon, which dominate the general-purpose server market, Vera is being positioned mainly for AI infrastructure, where CPUs help feed GPUs, run software orchestration, and handle supporting workloads around model training and inference. That makes early benchmarks important, but also harder to judge without broad workload coverage and power data.
Because Nvidia introduced Vera for AI datacenters in March 2026, the benchmark is not only relevant as a stand-alone server experiment. Oracle says it plans to deploy hundreds of thousands of Vera CPUs beginning in 2026, which makes the result relevant to a product Nvidia expects to ship at scale.
Michael Larabel, Phoronix reporter, spelled out Nvidia access limits before publishing the benchmark package.
“This isn’t a sponsored article but I obliged to their requests in order to run these initial Vera CPU benchmarks.”
Michael Larabel, Phoronix reporter (via Phoronix)
Larabel’s disclosure is the key caveat around the whole result. Nvidia framed the session around modern AI-datacenter workloads, but the company barred some of the validation points that enterprise buyers and competing vendors would expect to see.
“NVIDIA also requested only specific workloads relevant to the intended workloads/domains that Vera is catering to in the data center be tested. So this first round of Vera benchmarking isn’t too comprehensive across the spectrum of possible workloads but limited to the benchmarks that were permitted based on what they feel were most relevant — plus the fact I was only spending one day at NVIDIA’s offices. For these initial NVIDIA Vera benchmarks they preferred the scope of benchmarks be limited to target use-cases they feel most relevant for their modern data center customers.”
Michael Larabel, Phoronix reporter (via Phoronix)
What the first Vera test proved
Phoronix’s published benchmark mix covered code compilation, Python performance, OpenJDK Java workloads, AV1, and 7-Zip. Vera is not limited to a single demo, but the exposed basket still leaves open how the chip behaves in the wider enterprise tests Nvidia did not allow.
In the reported geomean results, Vera finished about 10% ahead of AMD EPYC and more than 50% ahead of Intel Xeon, with the reported geomean lead serving as the clearest summary number. Separate charts also put Vera close to AMD in video encoding while showing stronger per-core 7-Zip output.
From that mix, buyers can take two messages. Vera looks competitive in the workloads Nvidia exposed, but a narrower workload basket and missing power data still keep the result from serving as a full purchasing guide.
Vera also looks technically serious on paper. Its 88 in-house Olympus cores and 176 threads replace the off-the-shelf Arm core approach Nvidia used in Grace, while Armv9.2 compatibility and 1.2TB/s memory bandwidth give the chip the kind of data-feeding capacity AI systems need.
PCIe Gen 6 and CXL 3.1 support strengthen that positioning. Phoronix also listed a peak 450W TDP for the tested socket before memory draw, with LPDDR5X memory adding about 50W or less.
At the software layer, Nvidia’s published setup already reached Ubuntu 24.04 LTS with a patched Linux 6.18 LTS kernel, GCC 16.1, and driver support in Linux 7.1.
Those details narrow one common launch risk. Early server silicon often shows promising throughput before firmware, drivers, and platform tuning catch up, but Vera’s published environment suggests Nvidia has already pushed meaningful Linux support into place ahead of wider partner availability.
Vera’s AI-focused design and deployment stakes
Nvidia built Vera to remove CPU bottlenecks in tool-calling workloads and reinforcement-learning sandboxes that support larger AI systems. Ian Buck, a Nvidia hyperscale and HPC executive, has argued that fast CPUs keep code execution and sandbox work from slowing the GPUs around them, while Vera serves as the host processor for Vera Rubin NVL72.
Nvidia’s wider rack plan adds one more reason the benchmark matters. Rubin combines Vera with GPUs, networking, and DPUs in a rack-scale design, so CPU performance shapes how quickly the whole system can feed models, run tools, and manage orchestration work around the accelerators.
Competitive market and the validation gap
AMD EPYC and Intel Xeon remain the x86 server parts Vera is trying to displace in AI racks. Nvidia escalated Vera Rubin specs against AMD in January, and the new benchmark adds an early data point to that push.
Broader independent testing will decide how far that claim can go. Rubin systems are still slated for second-half 2026 partner availability, leaving time for outside testers to show whether Vera’s first public result holds up beyond Nvidia’s chosen workloads.
