AMD has published a fuller set of benchmark claims for its upcoming EPYC Venice server CPUs, including a result that puts its 256-core flagship well ahead of NVIDIA Vera in one SPEC CPU 2026 throughput test. The useful story is both the scale of AMD’s claim and the laboratory furniture around it: these are not clean, universal speed rankings.
AMD Watch analysis
What happened
According to Tom’s Hardware’s report, AMD’s new white paper expands its earlier EPYC Venice comparisons with more detailed results. AMD says a 256-core EPYC 9996 was 2.24 times faster than NVIDIA’s Vera in a SPEC CPU 2026 integer-rate throughput test, while a 96-core, high-frequency Venice configuration was around 20% faster in the cited comparison.
The same material puts the 9996 about 78% ahead of AMD’s previous-generation 192-core EPYC 9965. AMD also presents results for cloud, HPC and workloads labelled “agentic AI”, including NGINX, FAISS, TPC-derived tests and a replay of a multi-persona agent.
Key findings
- A huge headline number
AMD says the 256-core EPYC 9996 reached 2.24 times NVIDIA Vera’s score in one SPEC CPU 2026 throughput comparison. - A strong generational jump
AMD’s figures put Venice roughly 78% ahead of the previous-generation EPYC 9965 in the cited test. - Agentic AI gets a chart
AMD includes AI-adjacent server workloads, but some are derived or replayed tests rather than directly comparable industry benchmarks. - The comparison is not perfectly level
The configurations differ, and AMD used GCC 16.1 while NVIDIA’s Vera results used GCC 15.2.
Why it matters
Server buyers do not purchase a benchmark score in isolation. They purchase performance under a particular software stack, power limit, memory configuration, workload mix and price. Still, a claimed 2.24-times advantage is difficult to ignore when AMD is trying to establish Venice as a serious rival to NVIDIA’s expanding CPU-and-accelerator platform.
The AI angle is also revealing. AMD is framing general-purpose server CPUs as part of the infrastructure behind agentic workloads, not merely as background plumbing. That makes the comparison relevant beyond traditional databases and HPC, while also making careful test design more important. A workload called “agentic AI” is not automatically a portable, industry-standard measure of agent performance.
Our read
Venice looks technically significant, especially in AMD’s generational results, but the competitive headline remains an AMD claim built from mixed test conditions. Treat this as an important signal about AMD’s ambition and engineering progress, not a final league table.
The sensible reader move is to wait for independently controlled tests using matched compiler versions, power budgets, software and workloads. Server benchmarks are wonderfully good at producing large numbers and rather less good at choosing your procurement strategy for you.
What to watch
- Independent SPEC CPU 2026 results for production-relevant Venice configurations.
- Matched compiler, power and core-count comparisons with NVIDIA Vera.
- Real-world pricing, availability and performance-per-watt data.
- Whether AMD’s agentic-AI workload results hold up outside its own selected test suite.
Discussion spark: How much weight should buyers give AMD’s Venice headline gains before matched, independent tests and pricing are available?
Sources and evidence
Independent WittyWires tracker for public updates about AMD. Not affiliated with or endorsed by AMD; this is not an official account.