AMD's First EPYC Venice Benchmarks Take Aim at Nvidia Vera: What the 2.24x Claim Really Means

AMD has published its most detailed performance claims yet for its next-generation server processors, and the target is not only Intel. In a white paper released on Thursday, September 18, AMD pits its 6th Gen EPYC 9006 "Venice" chips directly against Nvidia's new Vera CPU.

The headline claim: in SPECrate 2026 Integer, AMD says its 256-core EPYC 9996 delivers 2.24 times the platform-level performance of an Nvidia Vera system and 1.2 times its per-core performance. Against Intel's flagship Xeon 6980P, AMD claims up to 2.37 times the throughput.

Those are big numbers. They are also vendor numbers, built on a mix of AMD's own testing, engineering estimates and competitor data gathered with different software. Here is what AMD is claiming, what the fine print says, and why the fight matters.

What AMD announced

The white paper evaluates Venice across general-purpose, enterprise, cloud-native, AI and high-performance computing workloads. In an AMD newsroom post, the company frames the paper around agentic AI, arguing that AI agents turn a single request into "a shifting chain of retrieval, tool calls, code execution and results," and that much of that work lands on CPUs rather than GPUs.

AMD's key claims from the paper and newsroom post:

• Versus Nvidia Vera: 2.24x platform performance and 1.2x per-core performance in SPECrate 2026 Integer, plus an estimated 3.4x throughput in a modeled 100-kilowatt rack.
• Versus Intel Xeon 6980P: up to 2.37x in SPECrate 2026 Integer, 2.4x to 3.7x in enterprise and cloud tests such as Redis, NGINX, MongoDB and OpenSSL, and 1.8x to 3.13x in molecular dynamics, materials modeling and weather forecasting.
• Versus its own last generation: Tom's Hardware reports AMD's slide shows the EPYC 9996 about 78% faster than the 192-core EPYC 9965.

AMD says Venice "is in production today," that major OEM platforms are on track to launch, and that leading cloud providers will begin deploying it later this year.

The three chips at a glance

Venice launched in July, and Vera and the Xeon 6980P are very different designs. Venice uses AMD's chiplet approach with compact Zen 6c cores for maximum density. Vera is a monolithic Arm-compatible chip that Nvidia pitches as the CPU built for AI agents. Intel's Xeon 6980P, from the Granite Rapids family, launched in September 2024 and is now two years old.

Spec sources: AMD and Tom's Hardware for Venice, Nvidia's Vera datasheet, and Intel's product listing for the Xeon 6980P. Theoretical bandwidth figures are calculated from channel count and memory speed.

The table makes the first caveat obvious. The EPYC 9996 has nearly three times the cores and threads of a Vera chip and draws considerably more power. A large platform-level win is expected. The interesting question is what happens when you control for that.

Reading the fine print

AMD's footnotes are unusually detailed, and they change how the headline numbers should be read. Tom's Hardware put it bluntly: "Mind the footnotes."

The 2.24x claim compares 512 cores with 176

According to AMD's newsroom footnotes, the 2.24x figure compares a two-socket EPYC 9996 server with an estimated SPECrate 2026 Integer score of 2,070 against a two-socket Vera system scoring 925. That is 512 Venice cores against 176 Vera cores.

The Vera number comes from Nvidia's own Vera white paper. The Venice number is labeled a "preliminary performance estimate based on AMD engineering projections." Neither is an audited SPEC submission.

The 1.2x per-core claim uses a different chip

The per-core claim does not come from the 256-core flagship. It comes from a 96-core high-frequency Venice configuration, which AMD estimates at 6.3 points per core versus 5.3 for Vera.

Run the flagship's own numbers and the picture flips. By AMD's figures, the two-socket EPYC 9996 works out to roughly 4.0 points per core, below Vera's 5.3. That is not a flaw in the chip; dense Zen 6c cores trade per-core speed for count. But it means "1.2x per-core" and "2.24x platform" describe two different processors.

Tom's Hardware also notes that for the detailed Vera comparison, AMD used a 256-core EPYC 9996 down-cored to 96 cores, and in earlier disclosures that configuration had the full 600 W power budget. AMD's actual 96-core high-frequency SKU tops out at 500 W.

Mismatched compilers

The white paper's breakdown of individual SPEC subtests compares Venice results compiled with GCC 16.1 against the Vera results Nvidia published using GCC 15.2. GCC 16 is the release that added Zen 6 support, so AMD's choice is understandable, but it folds compiler improvements into what looks like a silicon comparison.

"It's not best practice to compare benchmarks using two different compiler versions," Tom's Hardware wrote. Hardware Busters made a similar point, noting the comparison "folds architecture and compiler maturity into a single ratio with no clean way to separate them."

Memory bandwidth and the rack model

In the Stream memory bandwidth test, using data from Phoronix's earlier Vera testing, AMD says its down-cored 96-core configuration lands about 18% ahead of Vera overall and about 8% ahead per core. That is notable because Vera performed very strongly in Phoronix's original Stream results, but it is again a 96-core, 600 W configuration.

The 3.4x rack-level claim is modeled, not measured. AMD's footnote describes a geometric mean of six workloads within a 100-kilowatt power envelope, based on its own derivations and assumptions. TweakTown stressed that these are "modelled estimates" rather than "measurements from physical racks running both chips side by side."

Key takeaway: AMD's numbers show Venice is a very strong throughput chip, especially against its own previous generation. But the Vera comparisons mix estimates, competitor-published data, different compilers and non-shipping configurations. Treat them as a marketing opening bid, not a buying guide.

The claims, with caveats

Why the server CPU fight is heating up

For most of the past decade, the data center CPU story was AMD taking share from Intel. Nvidia's arrival as a standalone CPU vendor changes the shape of that contest.

Nvidia unveiled Vera as a standalone product on May 31 at GTC Taipei, saying Dell, HPE, Lenovo and Supermicro would offer it in standard single- and dual-socket servers, which Nvidia called "the first standard CPU option beyond x86." It named Anthropic, OpenAI and SpaceXAI among the AI labs planning to adopt it.

On its August 26 earnings call, Nvidia said Vera shipments were underway to lead partners including Oracle Cloud Infrastructure and SpaceXAI, with AWS starting in the current quarter. CFO Colette Kress said Nvidia expects its CPU revenue to "more than double" in fiscal 2028.

Why agents need CPUs

The reason both companies are shouting about agentic AI is that agents are CPU-hungry in ways chatbots were not. When an AI agent searches a database, calls a web API, spins up a sandbox to run code or checks its own output, that work typically runs on general-purpose CPU cores, not on the GPU generating tokens.

The two companies are betting on different answers:

1. Nvidia's bet is fast, efficient cores with very high memory bandwidth per core, tightly coupled to its GPUs. Nvidia says Vera provides up to 14 GB/s of memory bandwidth per core using low-power LPDDR5X.
2. AMD's bet is density and flexibility. It argues that a portfolio spanning 8-core edge chips, 256-core flagships and LPDDR-based "Verano" host processors lets customers pick the right chip for each stage of an agent pipeline on one software platform.
3. Intel's position is the most exposed. Its current flagship is two years old, and Tom's Hardware reports its next-generation Diamond Rapids data center CPUs are due next year.

There is also the question of what each chip is actually for. Vera is heavily optimized for sitting next to Nvidia GPUs and running orchestration work in AI factories. Venice is aimed at a much broader set of general-purpose server workloads. A head-to-head SPEC comparison is a useful talking point, but buyers are unlikely to choose between them purely on that basis.

What buyers should watch

If you are planning a server refresh for 2027, the white paper is a signal, not a spec sheet. Here is what will matter more:

1. Independent testing. Audited SPEC submissions and third-party reviews on shipping hardware, with matched compilers, will settle the per-core question.
2. Performance per watt and total cost. A 600 W chip and a 250 W to 450 W chip serve different rack budgets, and system pricing will matter as much as raw throughput.
3. Memory costs. Venice's 16-channel MRDIMM design delivers huge bandwidth, but server memory prices are elevated across the industry right now.
4. Software fit. Vera runs Arm binaries; Venice runs existing x86 software. For many enterprises, porting cost will outweigh a benchmark percentage.
5. Timing. AMD says Venice SP7 systems arrive in the fourth quarter, with the SP8, Venice-X and Verano variants following in 2027. Nvidia says Vera systems arrive from this fall.

FAQ

Is AMD EPYC Venice really twice as fast as Nvidia Vera?

In AMD's estimated SPECrate 2026 Integer results, a two-socket 256-core EPYC 9996 system scores 2.24 times a two-socket Vera system. But that compares 512 cores with 176, uses AMD engineering estimates and relies on Nvidia's published Vera score rather than side-by-side testing.

Is Venice faster than Vera per core?

AMD claims 1.2x per-core performance, but that figure comes from a 96-core high-frequency Venice configuration, not the 256-core flagship. By AMD's own numbers, the flagship's per-core score is lower than Vera's.

When can I buy EPYC Venice servers?

AMD says Venice is in production, OEM platforms are on track to launch and cloud providers will begin deploying it later this year. The first SP7 systems are scheduled for the fourth quarter of 2026.

How does Venice compare with Intel Xeon?

AMD claims up to 2.37x the SPECrate 2026 Integer throughput of the 128-core Xeon 6980P and 2.4x to 3.7x in enterprise and cloud workloads. The Xeon 6980P launched in 2024; Intel's next-generation Diamond Rapids chips are expected next year.

The bottom line

AMD's Venice white paper confirms the company has a formidable throughput monster on its hands. The generational gain over EPYC 9965 and the margins over Intel's aging Xeon flagship look genuinely impressive, even allowing for vendor optimism.

The Nvidia comparisons are more about positioning than proof. By mixing estimates, competitor-published data, mismatched compilers and non-shipping configurations, AMD has made a strong case for Venice's density while leaving the per-core question open. The real verdict will come when independent reviewers get both chips on the same bench, with the same software, later this year.