Leaked 256-core Zen 6c EPYC CPU destroys Zen 5 in early single-threaded benchmarks hints at 24% IPC gain

Semiconductors

Leaked 256-Core Zen 6c EPYC Shows 24% HTML5 Test Lead

An early Geekbench 7 comparison shows AMD's 256-core EPYC 9996 leading a clock-matched Ryzen 7 9700X, but cache and platform differences prevent a firm IPC verdict.

By
Tom Whitfield
Filed
Channel
Semiconductors
Read
4 min read

By Tom Whitfield

A leaked Geekbench 7 comparison placed AMD’s 256-core EPYC 9996 as much as 24% ahead of a clock-matched Ryzen 7 9700X in one Linux browser test. The result offers an early signal for Zen 6, but it does not establish the reported 15% gain in single-thread instructions per clock.

AMD’s EPYC 9006 server family has launched, although the source says the processors remain difficult for ordinary buyers to obtain and independent benchmarks are scarce. Tom’s Hardware contributor Zak described the early margin as “a massacre,” then cautioned that “there are just too many confounding factors to say that so far.”

What hardware did the comparison use?

The EPYC 9996 contains 256 Zen 6c cores arranged across eight “Monarch” CCDs, with 32 cores on each CCD. Every CCD carries 128 MB of L3 cache, divided into two 64 MB domains for each 16-core CCX.

That arrangement gives the server processor 1 GB of L3 cache and 4 MB per core, the same per-core ratio AMD has maintained since Zen 2. The Ryzen 7 9700X has eight Zen 5 cores and 32 MB of L3 cache, so its cache-per-core ratio matches, but its total local cache is far smaller.

The memory bus is 1,024 bits wide, and the source considers a 12,800 MT/s data rate likely; at that rate, bandwidth would reach 1.64 TB/s. The Ryzen 7 9700X test system used DDR5-6000 CL34 memory.

For single-threaded, lightly vectorized Geekbench work, neither memory system should be the main constraint. Matching both processors at 4.1 GHz improves the core-to-core comparison, but it does not remove every platform difference.

The takeaway is simple: the EPYC has far more cache and much greater memory bandwidth, making the performance lead hard to attribute to IPC alone.

What did the scores show?

Russian enthusiast “three twenty six,” also known as @326powah, ran the direct comparison after HXL, or @9550pro, spotted the original EPYC 9996 result. The server CPU ran inside a Dell PowerEdge R9815.

At 4.1 GHz, the EPYC 9996 recorded a single-thread score of 2,847. The Ryzen 7 9700X produced 2,483 under CachyOS Linux and 2,318 under Windows 11. Its stronger Linux score also shows how much operating-system differences matter.

In the Linux Gala-to-Honeycrisp results, the EPYC led across four workloads:

  • HTML5 Browser: 24% advantage
  • Video encoding: 21%
  • Asset compression: 19%
  • Photo editing: 17%

The browser result may partly reflect the EPYC’s 1 GB of L3 cache, because the HTML5 test is sensitive to memory latency. The other three gaps are less easily explained by configuration mismatch and point to architectural differences between Zen 5 and Zen 6.

The benchmark did not stress the new core’s SIMD capabilities. Geekbench uses relatively light SIMD work, and the test was marked AVX2, meaning it did not use the AVX-512 capability present in the processors. Much of the score therefore reflects integer arithmetic and memory behavior rather than maximum vector throughput.

The takeaway is that several workloads suggest real architectural gains, although one early comparison cannot establish a fixed IPC percentage.

Does the leak prove a 15% IPC gain?

No. Reports had already suggested a 15% single-thread IPC increase, and the new comparison makes that outcome plausible. It does not prove it, because cache capacity, operating systems, platform design and other variables differ across the two systems.

AMD designed Zen 6c as a high-density core for greater area efficiency, at the cost of clock scaling. The source calls 4.1 GHz modest for chips made on a bleeding-edge 2nm-class manufacturing process. Desktop Zen 6 processors will use a different core design, so the server result does not directly predict desktop performance.

Zen 5 delivered only a modest IPC increase over Zen 4, including some regressions, alongside a small clock-rate gain. That combination left many desktop users with little reason to upgrade. Zen 6 appears to offer a larger IPC lead in these tests, while rumors also point to a substantial clock increase over current parts. If both claims hold, the source compares the potential generational advance with the move from Core 2 to Core i7.

AMD has confirmed desktop Zen 6 processors for the AM5 socket and an early-next-year launch. The source gives no pricing or volume figures, and the author also hopes consumer RAM availability will keep pace. The takeaway is that the leak supports a Zen 6 lead, not a firm 15% claim for desktop chips. Early-next-year AM5 benchmarks will show whether Zen 6 can turn the server result into a broad competitive advantage.

Source: Tom's Hardware

Share this article:

More from Tom Whitfield

Tom Whitfield

Show full bio

Staff writer covering companies, deals and strategy in Semiconductors, chip manufacturing, hardware, AI compute at Chip Dispatch.

726 articles

Related articles

« Previous articleNext article »