October 23, 2026 — AMD's flagship server processor, the EPYC 9996, has surfaced in the Geekbench database: 256 cores and 512 threads, boosting to 4.12 GHz, scoring 2,847 single-core and 68,427 multi-core on Geekbench 7.1. Wccftech ran it within a day and IT Home followed on the 24th. This is the first public benchmark of the Zen 6 architecture — the EPYC 9006 "Venice" family that went on sale in July.

Here is the part most readers will get backwards: the raw single-core score is lower than a Zen 5 part. An EPYC 4545P scores 3,064 — because it boosts to roughly 5.5 GHz while the 9996 sample tops out at 4.12 GHz. Line the clocks up and Zen 6 is ahead by about 16.5% on average, with peaks of 24-25% in per-clock terms.

Read clocks before scores: single-core, correctedThe lower raw score is a clock gap, not a slower architectureEPYC 4545P (Zen 5, ~5.5 GHz)3,064 ptsEPYC 9996 (Zen 6, 4.12 GHz)2,847 pts9996 scaled to 5.5 GHz (est.)3,805 pts (+24%)The third bar is a linear clock scaling for illustration only, not a measurement. Sources: Geekbench 7.1 leak; Wccftech / IT Home.
Figure 1: Three steps for reading server scores: check the clock, scale it, then compare architectures. The third bar is an estimate.

Where the numbers come from

The data comes from the Geekbench database, reported by Wccftech on Oct 23 and rounded up by IT Home on Oct 24. The cleanest comparison did not come from AMD or a lab. A user with the handle @326powah locked a Ryzen 7 9700X at 4.1 GHz on DDR5-6000 CL34 memory and ran matched tests, which strips frequency out of the equation and leaves architecture. That is the basis for the 16.5% average and the 24-25% peak. Treat it as one careful data point, not a spec sheet: this is leaked data, on early parts, through a general-purpose benchmark.

How to read a leaked benchmark without fooling yourself

Three habits do most of the work. First, check what the score actually exercises: a general-purpose benchmark like Geekbench runs short bursts, not sustained all-core load under real memory pressure, and a 256-core part will never show you what it does to a consolidation ratio. Second, check the sample: public database entries can come from early silicon, partner labs or pre-production firmware, and this sample has not been confirmed by AMD. Third, check whether the comparison is fair. A 5.5 GHz versus 4.12 GHz clock gap is 33% — wide enough to swamp a 16.5% per-clock gain in the raw score. Anyone quoting 2,847 against 3,064 as an architecture verdict is reading the wrong column.

What survives all three filters is modest but real: at matched clocks and matched memory, Zen 6 does about a sixth more work per cycle than Zen 5. That is the number worth carrying into a procurement conversation.

Equal clocks: Zen 6 per-clock lead over Zen 5Ryzen 7 9700X locked at 4.1 GHz, DDR5-6000 CL34, matched runsAverage lead16.5%Peak lead25% (24-25%)Sources: Geekbench database leak; locked-clock comparison by @326powah; Wccftech (Oct 23) and IT Home (Oct 24).
Figure 2: Locked-clock testing removes frequency from the equation; what remains is the architectural gain.

EPYC 9996 at a glance

ItemSpecification
Architecture / familyZen 6, EPYC 9006 (Venice), launched July 2026
Cores / threads256C / 512T
Boost clock (leaked sample)4.12 GHz
Cache256 MB L2 + 1,024 MB L3
TDP600 W, configurable to 400 W
Geekbench 7.12,847 single-core / 68,427 multi-core
Reference priceUS$14,904 per 1,000 units for the 256-core flagship (AMD list; we cataloged all 31 SKUs previously)

Three takeaways

  1. This generation fights on IPC, not clocks. The counter-intuitive result — the flagship scoring below a mid-range Zen 5 part — is the design talking: the power budget feeds 256 cores and the memory subsystem, and frequency gives way. When you read server benchmarks, divide by clock first.
  2. 256-core chips are rarely limited by their cores. A Geekbench single-core score says little about consolidation ratios or database concurrency; those live on memory bandwidth and I/O. The 68,427 multi-core number proves the silicon runs; it does not promise your workload scales to it.
  3. For per-core licensed software, IPC is money. Databases and ERP systems that charge by the core get roughly one-sixth more done per license on a 16.5% IPC gain — same license budget, more effective throughput. Flip it around and discounted Zen 5 platforms need a price cut of about one-sixth just to match on work per license dollar — which is exactly when those discounts start appearing.

What a procurement team in the Pearl River Delta should do

Sources