AMD's first EPYC 'Venice' benchmarks aim straight at Nvidia

AMD has published the first official performance numbers for its EPYC 9006 server CPUs, the generation it calls Venice, and it has aimed them squarely at Nvidia's Vera CPU. The comparison sits in a new AMD white paper on the 9006 series, and the headline claim is a large one: in the SPECrate 2026 Integer suite a two-socket EPYC 9996 is estimated at 2,070 against an estimated 925 for a system built on an 88-core Nvidia Vera part, a ratio of 2.24x. AMD also puts the same EPYC part at 2.37x Intel's 128-core Xeon 6980P.

The platform underneath the claim

Venice is built on the Zen 6 and Zen 6c architectures and scales to 256 cores and 512 threads, with 16 channels of DDR5, second-generation MRDIMM support up to 12,800 MT/s, PCIe Gen 6, CXL 3.1 and up to 4TB of local memory per socket. AMD quotes about 1.6 TB/s of peak memory bandwidth per socket, against 1.2 TB/s for the LPDDR5X-based platform behind Vera.

Per-core is the number AMD actually wants quoted

The second comparison is the one aimed at anyone about to point out that a 256-core flagship meeting an 88-core part is not a fair fight. AMD estimates a 96-core EPYC 9006 processor at 1,210 in the same suite, or 6.3 per core, against 925 for Vera, or 5.3 per core, a 1.20x advantage, and says the 96-core part leads Vera in all 14 integer subtests. On memory bandwidth AMD measured 1,298 GB/s for a 96-core system in STREAM Triad against the 1,099 GB/s that Phoronix published for Vera, a 1.18x lead overall and 1.08x per core.

The fine print is doing real work

AMD is explicit that the SPEC figures are not published results. The white paper says results from both AMD and Nvidia were not yet suitable for publication under SPEC's rules on public system availability, so the numbers must be considered estimates. The Vera comparison data is taken from Nvidia's own architecture whitepaper and the STREAM comparison from a published Phoronix review rather than one matched test harness, which means two vendors' measurements are being placed side by side. The rest of the paper compares EPYC 9006 against Intel's Xeon 6980P and reports 2.4x for server-side Java, 2.5x for OpenSSL, 3.5x for MongoDB with YCSB, 2.9x for Redis, 3.7x for NGINX and 2.6x for MySQL transaction processing, plus HPC gains of 3.08x for NAMD, 1.80x for Quantum ESPRESSO, 3.13x for GROMACS and 2.90x for WRF. It models a 100 kW rack at 3.30x the platform performance of a Vera system. No prices, SKUs or availability dates were announced.

Our opinion

Read the white paper as a positioning document rather than a benchmark. AMD has chosen a comparison in which its 256-core flagship meets an 88-core competitor, then chosen a second comparison, per core on a 96-core part, for the audience that will immediately say the first one was unfair. Both live in the same document, which is more candour than this genre usually manages, and the footnote admitting the numbers are estimates is the most useful sentence in it. What is missing is the thing that decides a real purchase: a matched, reproducible result that neither vendor controls. SPEC's rule about public availability exists precisely to stop pre-launch silicon being crowned on the strength of a slide, and AMD concedes that is what these figures are. The wider shift matters more than the ratios. When the two most important CPU vendors in the AI build-out throw estimates at each other months before general availability, buyers are asked to plan capital spending around a spreadsheet. Use the tables as a list of workloads worth testing yourself, and nothing more.