What AMD actually claimed
AMD followed its July Venice launch with expanded performance claims in a new white paper. On the slide Tom's Hardware reviewed, the 256-core 9996 is 2.37x faster than the Intel Xeon 6980P and 2.24x faster than Vera. AMD's comparison with its own previous generation is also striking. Against the 192-core EPYC 9965, eTeknix reports the new flagship is 78% faster.
The benchmark is SPEC CPU 2026 Integer Rate. That test runs many copies of a workload on the same CPU, normally one copy per thread. A throughput test like this rewards whoever has the most cores. AMD's part has nearly three times Vera's core count, so a rate-based test favors it by design.
eWeek says AMD's own disclosure shows how much of the gap is core density. It lists an estimated SPECrate 2026 Integer score of 2,070 for a two-socket EPYC 9996 system and 925 for a two-socket Vera platform.
The fairer comparison: 96 cores versus 88
AMD also offers a closer matchup. It uses a 96-core Venice configuration that scores 1,210, or 6.3 per core, against 925 and 5.3 per core for Vera. That is roughly a 1.2x per-core advantage. That is where the "20% faster" figure comes from.
The caveats matter here:
- Down-cored silicon. The Vera figures come from NVIDIA's own white paper, and the 96-core part is an EPYC 9996 with several cores turned off.
- Power budget. AMD's white paper doesn't mention power. When AMD first shared the SPEC numbers, the 96-core model had the same 600W as the 256-core model, although the 96-core high-frequency SKU tops out at 500W.
- Compiler mismatch. In the Vera comparison AMD used GCC 16.1, while NVIDIA's published Vera figures used GCC 15.2. Different GCC major versions can shift results in either direction. AMD's reason, that Zen 6 needs GCC 16.1 support, is reasonable, but the test still isn't like-for-like.
- No official results. Neither AMD nor NVIDIA can publish official SPEC results yet, so all of the figures are estimates.
Tom's Hardware also reports a separate Stream memory test. In that test AMD showed about 18% more bandwidth, with the same 600W allowance on the down-cored chip. Keep that result apart from the SPEC numbers, because it is a different benchmark and configuration. AMD describes Venice as a 16-channel design with up to 12,800 MT/s RDIMMs, delivering 1,298 GB/s against Vera's 1,099 GB/s.
The rack-level claim is a model, not a measurement
AMD also claims about 3.3x to 3.4x throughput over Vera within a 100 kW rack budget. The sources report slightly different multiples. eWeek notes that the calculation covers six workloads and is not a measured head-to-head rack deployment. Another report says the estimates rely partly on modeled competitor performance rather than independently audited systems. Treat it as AMD's argument for density and efficiency, not as test data.
What it means for Windows and PC readers
For IT admins, Venice is relevant to future server refreshes, including virtualization and Windows Server hosts, but nothing here has been independently tested. The sensible move is to wait for matched-compiler results from third parties before putting it into a purchase plan.
For desktop users, the NoobFeed piece makes the key point: server chips differ from Ryzen in power limits, memory systems, cache layouts and core counts. Nobody should expect twice the gaming performance from these numbers. The one loosely transferable signal is that AMD is pitching per-core gains rather than only more cores, and even that comes from a down-cored, power-boosted server configuration.
NoobFeed also repeats rumors about 12-core desktop chiplets and a possible 24-core Ryzen. It says AMD has not confirmed these. It adds that AM5 support has been extended through 2029 and that BIOS updates will determine which boards accept future CPUs. I could not verify those platform claims independently, so check your motherboard vendor's CPU support list before buying.
Bottom line
- 2.24x is a 256-core versus 88-core throughput result in SPEC CPU 2026 Integer Rate. It is not a per-core or desktop result.
- About 1.2x per core comes from a down-cored 96-core chip with a mismatched compiler and a generous power allowance.
- 3.3x to 3.4x rack throughput is a modeled estimate, not a measurement.
- Independent, matched-toolchain testing is the missing piece. Until it exists, these are marketing-grade numbers.