AMD's EPYC Venice Benchmarks Reveal Nvidia's Real Threat

AMD wants you to look at one number: 2.24x. That is how much faster the company says its new EPYC "Venice" server chip is than Nvidia's Vera CPU. But the more interesting number is 20 percent, which is what is left of AMD's lead once you stop comparing chips with wildly different core counts. The real story in AMD's September 18 white paper is not a speed record. It is that AMD felt it needed to publish a chart at all, because Nvidia has quietly become a CPU competitor, not just the GPU supplier AMD's EPYC line has coexisted with for a decade.

The Headline Number Comes From Stacking More Cores, Not Faster Ones

AMD's flagship comparison pits a two-socket system built around its 256-core EPYC 9996 against a two-socket Nvidia Vera platform. That is 512 AMD cores against 176 Vera cores in the same power envelope, according to eWeek's read of AMD's disclosure. On AMD's estimated SPECrate 2026 Integer test, the EPYC system scored 2,070 against Vera's 925, a 2.24x edge that AMD has used as the headline of its own white paper and that outlets from Tom's Hardware to TechPowerUp have repeated. Do the division yourself: 2,070 divided by 925 is 2.238, so the multiplier checks out.

What the multiplier does not tell you is how much of that gap comes from AMD simply having roughly three times as many cores in the box. A chip with more cores will usually post a higher aggregate throughput score than a chip with fewer cores, almost regardless of architecture. That is not a flaw in AMD's math, but it does mean 2.24x is a statement about core density on this particular pair of systems, not a clean verdict on Zen 6 versus Nvidia's custom Arm design.

Match the Core Counts and AMD's Lead Shrinks to 20 Percent

AMD's own white paper anticipates this objection, which is why it also published a second, closer comparison: a Zen 6 configuration with 96 cores per socket against Vera's 88 cores per socket. There, AMD's per-core advantage falls to roughly 20 percent, or about 1.2 times Vera's per-core score, a figure Tom's Hardware, eWeek and wccftech all report independently. That is a real, plausible generational edge for a newer chip built on a newer process. It is also a much less dramatic story than "more than twice as fast," which is presumably why AMD led with the platform-level number instead.

AMD also models rack-level throughput under a 100-kilowatt power budget and lands at roughly 3.3x to 3.4x a comparable Vera rack, depending on which summary of the white paper you read, per Tom's Hardware and eWeek. That range itself is a tell: a modeled rack projection that shifts by a few tenths of a point depending on workload mix is a directional estimate, not a lab measurement.

The Compiler AMD Used Wasn't Even Available When Nvidia's Numbers Were Made

Hardware Busters flagged a detail that matters more than either headline multiplier: AMD compiled its own Venice results with GCC 16.1, the first compiler release with dedicated Zen 6 optimizations, while the Vera reference figures it compared against were built with the older GCC 15.2. A newer compiler tuned specifically for the chip under test can move benchmark scores on its own, independent of any silicon advantage. Hardware Busters called the result "a marketing number with an interesting engineering story underneath it," and that framing is fair: some of AMD's edge is real architecture, and some of it is AMD grading its own homework with a better calculator.

None of this makes AMD's numbers dishonest. The company's own methodology treats the figures as estimates rather than measured results, and outlets covering the white paper consistently describe them as AMD-modeled projections that have not been independently verified or SPEC-audited. That caveat belongs at the top of any story about them, not buried in a footnote, which is exactly where AMD put it.

Vera Isn't a Side Project, It's Nvidia's Bid to Own the Whole Rack

The reason AMD is benchmarking against Vera at all is the more important part of this story. Nvidia's Vera is an 88-core, Armv9-class CPU with two-way simultaneous multithreading for 176 threads, according to Nvidia's own Vera Rubin platform page. It is not a general-purpose server chip you can buy on its own. It is one of six components, alongside the Rubin GPU, an NVLink 6 switch, a ConnectX-9 SuperNIC, a BlueField-4 DPU and a Spectrum-6 Ethernet switch, that Nvidia sells as a single co-designed rack. Nvidia says Vera delivers roughly twice the performance of Grace, its first data center CPU, and the platform reached full production in June 2026 with partner availability arriving in the second half of the year.

That is the structural threat AMD is responding to. For most of the past decade, Nvidia needed AMD or Intel to supply the CPU next to its GPUs in an AI server. Vera Rubin removes that dependency for customers who buy the whole rack from Nvidia. AMD is not just defending EPYC's reputation for raw throughput here, it is defending EPYC's seat in racks that Nvidia would now rather fill entirely itself. That is a much higher-stakes fight than a spec sheet, and it explains why AMD reached for the most favorable multiplier it could support with a straight face.

Who Should Actually Care About This, and Who Can Ignore It

This fight matters to a narrow but consequential audience: engineers and buyers specifying CPUs for large AI training or inference clusters, cloud providers deciding how much of their next data center to buy as an integrated Nvidia rack versus a mixed AMD or Intel plus third-party GPU build, and investors trying to price how much CPU revenue Nvidia can pull away from AMD and Intel over the next few years. If you fall into any of those groups, the per-core number (about 20 percent) is the one to anchor on, not the platform-level 2.24x.

If you are choosing a desktop, workstation, or a small business server, this entire comparison is irrelevant to you. Vera does not ship as a standalone chip you can buy, EPYC Venice server parts are not yet broadly available, and neither product competes in the consumer or SMB space where GPU choice or a standard x86 workstation chip is what actually decides your purchase.

The Honest Case for Taking AMD's Numbers Seriously Anyway

The skeptical read above is fair, but it would be wrong to wave the whole white paper away as pure marketing. AMD is not claiming parity with Nvidia on paper specs, it is claiming a real per-core architectural gain of about 20 percent, a number modest enough that a company willing to inflate results would have picked a bigger one. Zen 6's move to 12-core CCDs, up from eight-core CCDs in Zen 5, plus 16 memory channels and PCIe 6.0, are genuine platform upgrades independent of any benchmark, as ServeTheHome's coverage of AMD's architectural disclosure lays out. A 20 percent per-core gain on a next-generation chip against a first-generation custom Arm design from a GPU company is a plausible, even conservative-sounding claim, not an obviously cooked one.

It is also worth remembering that Nvidia has never published independent, audited numbers proving Vera beats EPYC either. Both companies are running the same playbook of self-serving benchmarks ahead of third-party testing, and AMD going first does not make Nvidia's silence more credible.

What to Watch Before You Trust Either Company's Numbers

Use this checklist before you let any vendor benchmark chart change a buying decision or an investment thesis:

• Check whether the comparison uses matched core counts or matched sockets. A platform-level multiplier with a large core-count gap tells you about density, not architecture.
• Check the compiler and software stack on both sides. A newer compiler tuned for one chip can move scores by double digits on its own.
• Look for the word "estimated," "modeled," or "projected." AMD's own figures carry that label, and so will most first-party chip benchmarks you will see this year.
• Wait for a third party. SPEC audits, Phoronix, ServeTheHome and Tom's Hardware typically publish independent numbers once hardware ships in volume, usually months after a vendor white paper.

Choose to weight AMD's claims heavily only if your workload maps closely to SPECrate integer throughput and you can wait for independent verification before committing budget. Skip weighting them at all if you are buying anything smaller than a rack-scale AI cluster, since neither chip is relevant at that scale.

The Takeaway

AMD's white paper is less a benchmark story than a positioning story. The 2.24x figure is real arithmetic on real, if AMD-modeled, numbers, but it measures core count more than silicon. The 20 percent per-core figure is the more honest claim, and it is the one to watch as both companies push toward rack-scale AI hardware, alongside AMD's broader bet on AI infrastructure customers and other chipmakers' moves against Nvidia. Wait for independently audited numbers from a third party like Phoronix or SPEC before treating either chart as settled fact.

Sources

• Tom's Hardware: AMD shares first official benchmarks for EPYC Venice CPUs, targets Nvidia
• Tom's Hardware: AMD fires back at Nvidia, claiming 256-core Zen 6 Venice CPU beats Vera by 3.3x in rack-level performance
• eWeek: AMD Says EPYC Venice Beats Nvidia Vera by 2.24x, But Its Own Data Shows a Much Closer Race
• Hardware Busters: AMD's first official EPYC Venice benchmarks aim straight at Nvidia Vera, and the compilers don't match
• wccftech: AMD targets Nvidia Vera in latest EPYC Venice agentic AI benchmarks
• Nvidia: Vera Rubin Platform overview
• ServeTheHome: AMD takes the lid off next-gen EPYC 9006 Venice as Zen 6 comes to servers