NVIDIA Vera Challenges Xeon and EPYC: Initial Benchmarks Show the Potential of the Arm CPU with Olympus Cores

Phoronix has published the first performance proofs of the NVIDIA Vera CPU, an Arm-based solution specifically intended for use in datacenters for AI workloads and more. The initial results are particularly promising.

Michael Larabel, founder of the site specializing in Linux, described Vera as "the fastest Arm processor on Linux ever tested in the history of the publication," which has been active for over twenty years. However, it should be noted that the benchmarks were conducted directly at NVIDIA's headquarters in Santa Clara and in a context strongly controlled by the company.

This means the results should be interpreted with caution: the processor is still in pre-production, and NVIDIA has limited both the allowable tests and the available monitoring tools, including detailed measurements of power consumption.

Vera represents a significant change in NVIDIA's CPU strategy. Unlike the previous Grace solution, which was based on Arm-designed Neoverse cores, the new solution uses proprietary "Olympus" cores developed internally.

The CPU features 88 Armv9.2 cores with 176 threads and, among the most relevant features, includes native support for FP8 instructions, increasingly utilized in modern AI workloads. NVIDIA also implements an SVE2 6x128-bit configuration designed to accelerate certain workloads directly on the CPU without completely offloading to the GPU.

The platform provides 1.2 TB/s of memory bandwidth and supports up to 1.5 TB of LPDDR5X memory through SOCAMM2 modules. NVIDIA has also developed a second generation of its Scalable Coherency Fabric interconnect, capable of delivering 3.4 TB/s of bandwidth within a unified monolithic die. The aim is to reduce the latency typically associated with chiplet architectures adopted by many competitors.

Phoronix compared Vera with several recent server CPUs, including Intel Xeon "Granite Rapids" 6980P and various AMD EPYC "Turin" and "Turin Dense" models, including EPYC 9755, 9575F, and 9475F. The previous NVIDIA Grace CPU was also included in the tests.

According to the geometric mean of the authorized benchmarks, Vera achieved performance superior by about 11% compared to the best tested AMD solutions and over 55% compared to the best single-socket Xeon considered. In some cases, the NVIDIA CPU managed to outperform even dual-socket configurations, highlighting how certain workloads do not scale perfectly on multi-processor platforms.

The most evident advantages emerge in stream testing and 7-Zip compression, while in other scenarios, high-end EPYC models manage to stay very close to the NVIDIA offering.

The authorized benchmarks included software compilation, memory performance, video encoding, Python and Java workloads, databases, and synthetic workloads.

It was not possible to take reliable measures of energy efficiency, one of the main strengths traditionally associated with Arm architectures. The declared TDP for Vera is 450 W, plus about 50 W for the 768 GB memory pool used in the tested configuration.

With Vera, NVIDIA seems to want to take on a much broader role in the infrastructure market. According to reported estimates, the company could generate about $20 billion from the Grace and Vera platforms, entering an overall addressable market estimated at around $200 billion.

The company is already collaborating with major hyperscalers for the integration of rack-based solutions on Vera, both for internal use and for cloud offerings aimed at third parties. This approach could allow NVIDIA to quickly become one of the leading datacenter CPU manufacturers.