AMD's case against Nvidia Vera, as told by its footnotes

On September 18 AMD put out a newsroom post and a 34-page white paper, AMD EPYC Server CPU Architecture and Performance Overview, making its first detailed performance case for the 6th Gen EPYC 9006 "Venice" family. Nvidia's Vera CPU is the main target. The newsroom post's headline sentence is this:

In SPECrate® 2026 Integer testing, the AMD EPYC™ 9996 server CPU delivers 1.2 times the per-core performance of an Nvidia Vera-based platform and 2.24 times its platform-level performance.

AMD newsroom, September 18, 2026

Both numbers are in the white paper. The endnotes, where the configurations are listed, tell a different story about what was compared.

Measured, or projected?

The white paper's own status for these numbers is not consistent. The body text, introducing the SPEC CPU 2026 results, says: "As this document is published, results from both AMD and Nvidia were not yet suitable for publishing based on SPEC rules regarding public system availability. The results discussed below are measured results that must be considered estimates based on publishing rules."

The footnotes behind those same figures (endnotes 3, 4 and 7) each say something else: "Preliminary performance estimates based on AMD engineering projections and subject to change." Endnote 6 says "based on AMD internal testing." A reader cannot tell from the document which numbers came off a machine and which came out of a model.

The 2.24x: platform against platform

Endnote 3 gives the inputs. A 2P EPYC 9996 system has an "estimated SPECrate®2026_int_base score 2070"; Nvidia Vera, "88 cores," has an "estimated SPECrate®2026_int_base score 925." Both are listed with GCC 15.2. 2070 divided by 925 is 2.24.

The 9996 is a 256-core part, so the 2P system has 512 cores. The Vera figure is described as 2P in the same note, which means 176 cores. Divide each score by its cores and, on AMD's own numbers, the 9996 system produces about 4.04 per core against Vera's 5.26. That makes the chip in the headline about 0.77x of Vera per core. The 2.24x is a density result: more cores per system. It is not a faster core.

The 1.2x: which 96-core part?

The newsroom sentence gives the 1.2x per-core figure to "the AMD EPYC™ 9996." Read on its own, that means the 256-core chip from the 2.24x claim. The endnotes describe the tested part three different ways.

Endnote 4 (claim 9xx6-049) credits a "2P 6th Gen AMD EPYC 96c HF processor, GCC 15.2, estimated SPECrate®2026_int_base score 1210 (6.3 per core)," against Vera at "925 (5.3 per core)", also GCC 15.2.

Endnote 7 (9xx6-053) is the per-subtest table that the body text says "offers a more detailed view into the results in Figure 10." It describes the AMD system as a "6th Gen AMD EPYC 96-core CPU (downcored from 256C): GCC 16.1, MRDIMM-8000," and the Vera side as "scores from NVIDIA published architecture whitepaper." Its geomean row reads 1210.31 924.87 1.20x. Endnote 8, on memory bandwidth, calls the same kind of system "1P 6th Gen AMD EPYC 9996 (96-core)."

So the "9996" in the per-core claim is either a separate 96-core high-frequency part or a 256-core 9996 with 160 of its cores switched off, depending on which footnote you trust. The same 1210 also appears under two compiler versions. In the detailed table, one side is AMD's own run on GCC 16.1 and the other is copied from Nvidia's document. Endnote 6 is built the same way: it covers the four subtests Nvidia had highlighted (cpython, gcc, llvm, cppcheck), and there the per-core margins are 1.12x, 1.01x, 1.07x and 1.04x.

The older SPECrate 2017 per-core comparison in endnote 5 mixes toolchains openly: "6th Gen AMD EPYC (96C and 128C) using AOCC 5.1 ... NVIDIA Vera using GCC 13."

The rack number moves between pages

For the 100 kW rack model, the white paper body says "6th Gen AMD EPYC™ 9996 Server CPUs extend this comparison to 3.30x of Nvidia Vera-powered platform performance" and puts the 5th Gen 9965 at "an estimated 2.37x." It adds that the estimates use a geomean "of six workloads including SPECrate® 2017_int_base."

Endnote 27, the source for that figure, lists the six workloads as including "SPECrate®2026_int_base," not 2017. Its results table reads "5th Gen AMD EPYC 9965 192C 2.30" and "6th Gen AMD EPYC 9996 256C 3.40." The newsroom post uses the footnote's version: "an estimated 3.4 times the throughput of a Vera-based platform." The rack model itself is spelled out: a Dell IR7000, Vera at 450W with 176 CPUs per rack, the 9996 at 600W with 126.

How to read it

None of this says Venice is slow. On these figures a 96-core Venice part does beat an 88-core Vera per core, by 1% to 101% depending on the subtest (stockfish is the 2.01x outlier). And a 512-core 2P box outruns a 176-core one. Those are real positions to take.

What the document does not do is keep its own inputs straight. The same headline number is labeled measured and projected. The per-core claim carries the name of a 256-core chip but was measured, or projected, on 96 cores. The per-subtest table compares an in-house run on a newer compiler with a competitor's published figures. And the rack estimate is given once as 3.30 and once as 3.40. AMD says "'Venice' is in production today." By AMD's own account, SPEC's public-availability rules are what kept both vendors' results unpublished. Until audited results appear, the endnotes are the most useful part of this document.

Primary sources: AMD white paper, "AMD EPYC Server CPU Architecture and Performance Overview" (body text and endnotes 3-7, 27), AMD newsroom, "EPYC CPUs Deliver for Every Layer of the Agentic AI Stack", read 2026-09-25.