AMD Claims 256-Core Zen 6 EPYC CPU Smokes NVIDIA Vera In AI Servers

A couple of weeks ago, NVIDIA gave Linux blog Phoronix exclusive access to a Vera system to run a few benchmarks on it. Vera is of course NVIDIA's next-generation CPU intended to be paired with its Rubin GPUs, and it features the company's in-house "Olympus" CPU cores, making it something of a wildcard in the market. However, the benchmarks that Phoronix was allowed to run were tightly constrained by NVIDIA, and the comparisons that the site was allowed to make were obviously crafted to highlight Vera's strengths. Now, AMD is making its own comparisons that put its next-gen EPYC CPUs in a much better light, but its numbers are specifically curated as well.

In a blog post titled "Agentic AI Needs Rack-Scale CPU Performance – AMD EPYC Delivers It Today," AMD argues that thread-, core-, or even socket-level performance comparisons aren't really meaningful because they don't describe what a customer actually deploys. This part of the argument is reasonable, but AMD then goes on to admit that almost all of the data in its post is actually modeled, not actual benchmarks. That makes these comparisons plausible, but not real-world. Nonetheless, the data is interesting as it gives us our first look at what AMD expects to achieve with its next-generation "Venice" processors based on Zen 6.


Based on a rack-level comparison using modeled data, AMD claims that its extant 192-core Turin processors already deliver nearly two and a half times the performance per watt of NVIDIA's Vera. (Interestingly, the company also notes that even Intel's Granite Rapids is almost half again as fast as Vera.) Then, the new EPYC "Venice" part with 256 cores will apparently increase that lead all the way to 330%.

AMD's language in the post doesn't pull any punches. The house that Zen built says "the picture is clear: The density positioned as future-looking is already being exceeded with standard infrastructure available now." Harsh words for NVIDIA's new server CPU, which in Phoronix's directed testing was slightly faster than AMD's extant Zen 5-based parts in many benchmarks. However, those comparisons used a 64-core EPYC processor, not the top-end 192-core dense (Zen 5C) model nor even the fastest 128-core standard-density chip, featuring higher clocks.

Clocks are important, because threaded performance is important. AMD says that its 64-core Venice chips are projected to be 27% faster than the 88-core Vera on a per-core basis, and that the 96-core Venice will still be 11% faster per-core despite offering more cores.


"Projected" is an important word there, though. In its Methodology section, the company says that "Comparisons are performed at rack level using a reference 100 kW envelope with 2P platforms, with system power and nodes-per-rack normalized to NVIDIA Vera. Because the 'Venice' and Vera figures reflect modeled and projected configurations, results are presented as estimates within the stated rack-power constraint."

In other words, there's minimal data here for the interesting chips; the relationship between Turin and Granite Rapids is well-understood, while Vera and Venice are completely based on AMD's models. Furthermore, while AMD says "Agentic AI," the benchmarks run are SPEC CPU, SPECjbb Java server benchmark, NGINX web hosting, redis-benchmark, Memcached, and TPROC-C MySQL database benchmark. Of these, only redis-benchmark is really even related to AI performance; the rest are more standard data center and HPC benchmarks.

We're not saying that the data is invalid, though. If anyone would know how Venice will perform, it would be AMD. Of course, there's a huge incentive for AMD to portray its own product in the best possible light and the competitor's product in less favorable light. Understanding that, we have to take these results with a grain of salt.

Indeed, arguably the most convincing point that AMD makes is that its chips are tried-and-true x86 processors on which you can run any x86-64 compatible code, which makes up the vast overwhelming majority of server software these days. Vera is based on an Arm architecture, which has very small benefits in terms of energy efficiency, but major downsides in terms of software compatibility.