AMD has published its first detailed benchmark white paper for EPYC Venice, claiming the 256-core flagship is 2.24x faster than NVIDIA's 88-core Vera in SPEC CPU 2026 integer throughput. The configuration and compiler differences matter more than the headline multiples.
What was announced
- AMD published an EPYC server CPU architecture and performance white paper on Friday 18 September 2026, extending the claims made when Venice launched in July 2026.
- In SPEC CPU 2026 intrate, AMD puts the 256-core EPYC 9996 at 2.37x the Intel Xeon 6980P and 2.24x NVIDIA's 88-core Vera.
- AMD claims the 9996 is about 78 percent faster than the previous-generation 192-core EPYC 9965 in the same test.
- AMD maintains that a 96-core high-frequency Venice part is about 20 percent faster per core than Vera in SPEC CPU 2026 integer rate.
- The SPEC throughput numbers were gathered in July 2026 with GCC 15.2; the per-core subtest comparison used GCC 16.1 against NVIDIA's own Vera figures, which were produced with GCC 15.2.
- The per-core comparison used a down-cored EPYC 9996 running 96 of its 256 cores, with no power budget stated; AMD's 96-core high-frequency Venice SKU tops out at 500 W against 600 W for the 256-core part.
- Other workloads shown include NGINX, the TPCx-AI kit, FAISS, workloads derived from TPC-H and TPC-C, and a multi-persona agent replay.
The ETON view
Two vendors comparing their own CPUs with different compilers is a marketing exercise, not a procurement input. What is useful here is the shape of the claim: AMD is arguing that agentic AI workloads, retrieval, web serving, databases and the orchestration around a GPU are won on core count and per-core throughput, which is the part of an AI platform that is easiest to buy sensibly.
For anyone planning a 2026 to 2027 refresh, the decision is rarely Venice against Vera. It is whether to wait for a part that is still pre-launch on a vendor-controlled benchmark, or to buy EPYC 9004 and 9005 or Xeon 6 capacity now at a known price with a known lead time. Down-cored, high-frequency SKUs are also worth a second look: if your licensing is per core, a 96-core part that clocks higher can cost less to run in software terms than a 256-core part you never saturate.
We are vendor neutral on this, which means we will quote the same workload three ways: current-generation new, previous-generation refurbished with warranty, and a mixed estate where the GPU nodes are new and the supporting compute, storage and networking are not. Bring us the workload rather than the part number and we will tell you where the money actually goes.
Related infrastructure
Category: Servers · Vendor: AMD, NVIDIA, Intel · Technology: EPYC Venice (Zen 6), NVIDIA Vera, Xeon 6980P, SPEC CPU 2026 · Last verified: 20 Sep 2026 · ~2 min read
Sourcing this kind of infrastructure? Talk to ETON about availability, lead time and pricing.
