Early benchmark data for AMD’s Zen 6 core architecture is pointing to a generational step rather than a modest tune-up. A 256-core EPYC 9996 processor has surfaced in the Geekbench database with a single-core score of 2,847 — a figure that, once frequency differences are accounted for, implies roughly 24 percent more work per clock than comparable Zen 5 silicon.
- Key Point: The EPYC 9996 pairs 256 cores and 512 threads with 1 GB of L3 cache, 256 MB of L2 and a 600 W thermal envelope that can be configured down to 400 W.
- Key Point: Its 2,847 single-core result came at a 4.12 GHz ceiling, while a 16-core EPYC 4545P boosting to 5.47 GHz scores 3,064 — a gap that works out to about 24% more performance per clock for Zen 6.
- Key Point: An enthusiast running a Ryzen 7 9700X pinned at 4.1 GHz with DDR5-6000 CL34 memory recorded 2,444 points, putting the Zen 6 part 16.5% ahead on average and up to 24-25% ahead in some subtests.
- Key Point: Desktop Zen 6 processors, reportedly codenamed Olympic Ridge, are expected in 2027 on the existing AM5 socket with up to 24 cores and up to 96 MB of L3 cache.
A flagship server part sets the baseline
The EPYC 9006 “Venice” family began shipping in volume several months ago, and retail silicon is now turning up in public databases. The 9996 sits at the top of that stack: 256 cores, 512 threads, 1 GB of L3 alongside 256 MB of L2, and a 600 W rating that partners can dial back to 400 W. Its Geekbench 7.1.0 run produced 2,847 points in the single-core test and 68,427 across all threads, with the core clocked at 4.12 GHz — also its maximum boost frequency.
Taken alone, 2,847 looks unremarkable next to existing Zen 5 server parts. The number only becomes interesting when it is set against faster-clocked hardware. A 16-core EPYC 4545P reaching 5.47 GHz scores 3,064 points, so the 9996 delivers close to the same per-core result while running more than 1.3 GHz slower. Divide the scores by frequency and Zen 6 comes out as much as 24% ahead per clock.
A second data point comes from consumer hardware rather than a datacenter part. One enthusiast locked a Ryzen 7 9700X at 4.1 GHz with DDR5-6000 CL34 memory and logged 2,444 points. At an essentially identical clock, the Zen 6 chip finished 16.5% ahead on average, with individual subtests opening a gap of 24% to 25%. Because one run came from a server platform and the other from a desktop setup, the pattern suggests the uplift belongs to the core itself rather than to a particular operating system or memory subsystem.
Compact cores and a single synthetic test
Venice processors use Zen 6C, a variant of the architecture built for denser placement and better energy efficiency. It shares the same instruction set as the standard core, but its compact layout means the server submission is not a perfect proxy for what a full-size desktop Zen 6 core will do.
The results are preliminary: Geekbench is one synthetic workload, comparisons across platforms are approximate, and neither score has been independently verified.
What the numbers suggest for desktop Ryzen
The Olympic Ridge family is expected to arrive in 2027 on AM5, keeping the socket that existing Ryzen owners already have. Reported specifications point to a doubling of the core ceiling per chiplet, from eight cores on Zen 3 through Zen 5 to twelve on Zen 6, which lifts the maximum core count from 16 to 24 and the L3 pool from 64 MB to 96 MB across the package.
Desktop parts should also differ from the server chip in ways that favor performance: full-size Zen 6 cores instead of compact ones, higher boost clocks, and faster memory on enthusiast boards. If the per-clock advantage holds, the eventual desktop gains could exceed what the EPYC submission shows. Intel is expected to field Nova Lake-S in the same 2027 window on TSMC’s N2P process, so the two lineups appear set to land on comparable manufacturing technology.
For buyers, the significance of a per-clock improvement is that it compounds. Higher instructions-per-cycle figures raise performance in every workload, including the single-threaded tasks that still dominate gaming and everyday responsiveness, and they do so without the power cost of pushing frequencies higher. That is the metric to watch as more samples reach public databases. The caveat is that a 256-core server part on a compact core design is a long way from a retail Ryzen chip, and AMD has not confirmed clocks, core counts or pricing for Olympic Ridge. Until desktop silicon is tested, the 24% figure should be treated as a promising ceiling rather than a promise.