Leaked EPYC 9996 Benchmark Puts Zen 6c 24% Ahead of Zen 5

A single-threaded Geekbench 7 score of 2847 — that's what a leaked run of AMD's 256-core EPYC 9996 posted on a Dell PowerEdge R9815, beating a Ryzen 7 9700X by up to 24% in matched-clock testing. The comparison comes from Russian enthusiast "three twenty six" (@326powah), who pitted the server chip against his own desktop part at an identical 4.1 GHz. Well-known leaker HXL (@9550pro) spotted the original EPYC 9996 result first.

The headline number everyone's repeating — a 15% IPC uplift for Zen 6 over Zen 5 — is plausible but not proven. Too many confounding factors sit between these two machines to call it settled. Here's what actually ran, and what it changes for practitioners watching AMD's roadmap.

What hardware was tested?

The EPYC 9996 is a gargantuan part: 256 Zen 6c cores spread across eight "Monarch" CCDs, 32 cores apiece. Each CCD carries 128MB of L3 cache, reportedly split into two 64MB domains per 16-core CCX. That works out to 4MB of L3 per core — the exact ratio AMD has kept since Zen 2 — and 1GB of L3 on the package overall.

The comparison machine, a Ryzen 7 9700X, has eight Zen 5 cores and 32MB of L3. Same cache-per-core ratio, much smaller local pool. The EPYC also rides a 1024-bit memory bus, likely at 12,800 MT/s, for a colossal 1.64 TB/s of bandwidth. For single-threaded Geekbench workloads that barely matters — even 326's DDR5-6000 CL34 desktop memory isn't the bottleneck.

Matching both chips at 4.1 GHz is what makes the dataset interesting. It's a reasonably clean core-to-core comparison, even if nothing else is equal.

How big are the gains?

It's a massacre on both operating systems. The EPYC 9996 scored 2847 single-threaded; the 9700X managed 2483 on Linux (CachyOS) and 2318 on Windows 11. Zen 5 takes its beating more gracefully on Linux, but it still loses everywhere.

Looking at the Linux results test by test:

• 24% in the HTML5 Browser test — sensitive to memory latency, so the EPYC's 1GB L3 could explain much of it
• 21% in video encoding
• 19% in asset compression
• 17% in photo editing

The last three can't be waved away as configuration mismatches. They point to genuine architectural differences between Zen 5 and Zen 6.

Why the SIMD caveat matters

Here's the fascinating part: these results don't lean on Zen 6's SIMD muscle. Geekbench historically makes light use of vector units, and this specific test is marked "AVX2" — meaning it never touches the AVX-512 units present in both CPUs. Most of these subtests measure integer arithmetic and memory subsystem behavior rather than raw number-crunching. Zen 6 winning this decisively is genuinely interesting for that reason.

What does this mean for desktop Zen 6?

Don't treat this as gospel on IPC, and remember the core in question isn't the one you'll buy. The EPYC 9996 uses Zen 6c, the dense variant built for area efficiency at the cost of clock scaling — 4.1 GHz isn't exactly ambitious on a 2nm-class process.

That's actually the encouraging part. Zen 5 delivered only a modest IPC uplift over Zen 4, with some regressions, plus similarly modest clock gains — a flat release for desktop buyers. Zen 6 now looks likely to pair a significant IPC advantage with, if rumors hold, a fairly extreme clock jump over current parts. If that lands, this could be a generational leap on the scale of Core 2 to Core i7.

AMD has confirmed these parts are coming to the AM5 socket. Launch is expected early next year — with any luck, alongside RAM you can actually afford to buy.