Leaked 256-core Zen 6c EPYC hints at a big IPC jump

AMD’s next server processor has turned up in Geekbench ahead of any official announcement, and the single-core result is good enough that desktop builders are paying attention to a chip they’ll never buy. The EPYC 9996, a 256-core part built from dense “Zen 6c” cores, reportedly scored 2,847 points in the Geekbench 7.1 single-core test at around 4.1GHz. Everything here is leaked, unofficial and almost certainly from engineering-stage hardware, so read it as a rumour with numbers attached.

What the leak actually shows

The result was spotted by the enthusiast HXL (@9550pro) on X, and Tom’s Hardware reports that the run came from a Dell PowerEdge R9815 server. A second enthusiast, “three twenty six” (@326powah), then pitted it against his own Ryzen 7 9700X locked to the same 4.1GHz, which gives a rare clock-matched look at one generation of AMD cores against the next. VideoCardz relayed the comparison on 24 October.

The headline figure depends on whose baseline you use. VideoCardz puts the Ryzen 7 9700X at 2,444 points, a gap of 16.5% in the EPYC’s favour. Tom’s Hardware quotes 2,483 points for the Ryzen under Linux and 2,318 under Windows, and its headline floats a 24% IPC gain, a number that comes from a different comparison (more on that below).

Individual subtests vary a lot. VideoCardz lists gains of up to 25.7% in single tests, while Fudzilla, citing Guru3D, puts the spread at 8.2% to 25.7%, with HTML5 Browser, Photo Editor, Structure from Motion and Video Player among the bigger jumps. File Compression and Clang moved least. Tom’s Hardware notes there are no regressions in the Linux results, which is more than Zen 5 managed against Zen 4.

EPYC 9996 specs, as far as they’re known

The 9996 is the top of AMD’s upcoming “Venice” family. Tom’s Hardware describes it as 256 Zen 6c cores spread over eight CCDs of 32 cores each, with 128MB of L3 cache per CCD, which adds up to 1GB. That keeps AMD’s long-running ratio of 4MB of L3 per core.

Memory is the other eye-catching number. Tom’s Hardware says the chip has a 1024-bit memory bus, likely running at 12,800 MT/s, for 1.64 TB/s of bandwidth; treat that as an informed estimate rather than a confirmed figure. Wikipedia’s entry for the EPYC line lists the Venice family as the first Zen 6 product, built on a TSMC 2nm-class node, with 16 DDR5 channels, 128 lanes of PCIe 6.0 and TDPs up to 600W across the series. AMD hasn’t published a TDP for this particular SKU.

Spec	EPYC 9996 (leak)	EPYC 9965	Xeon 6990E+
Core type	Zen 6c	Zen 5c	Darkmont E-core
Cores / threads	256 / 512	192 / 384	288 / 288
L3 cache (MB)	1,024	384	576
Clock, top (GHz)	~4.1 (sample)	3.7	3.2
Memory channels	16 (series spec)	12	12
Memory speed (MT/s)	~12,800 (estimate)	6,400	8,000
TDP (W)	Not reported	500	450
Process	2nm-class	3nm	Intel 18A

The Xeon column assumes one thread per core, since Intel’s Clearwater Forest cores don’t use Hyper-Threading. AMD’s Zen 6c cores reportedly do, which is why the EPYC lands at 512 threads.

How Zen 6c stacks up against Zen 5

The cleanest comparison is the clock-matched one: 2,847 points for the EPYC 9996 against 2,444 for a Ryzen 7 9700X, both near 4.1GHz. A second comparison uses the Zen 5-based EPYC 4545P, a 16-core part with a boost clock up to 5.5GHz, which scored 3,064 points at about 5.45GHz.

Chip	Core	Clock (GHz)	Geekbench 7 single-core (points)
EPYC 9996 (leak)	Zen 6c	~4.1	2,847
Ryzen 7 9700X	Zen 5	~4.1	2,444
EPYC 4545P	Zen 5	~5.45	3,064

Geekbench 7 single-core score (leaked results)

The older chip keeps the higher absolute score, but it needs a clock roughly 33% higher to do it. The Zen 5 part is only 7.6% ahead of the leaked Zen 6c result, according to Overclock3D, which calculates an IPC gain of about 23% once the frequencies are normalised. That’s the source of the 24% figure in Tom’s Hardware’s headline.

Two caveats come with it. The 4545P comparison spans two different server platforms, and a French outlet, Hardware Cooking, points out that it doesn’t isolate IPC cleanly, so the 16.5% clock-matched figure is the safer one to quote. And a dense-core server chip, a Dell server and a desktop board with DDR5-6000 CL34 memory are not identical test benches, even if single-threaded Geekbench is unlikely to be limited by memory bandwidth.

Why the numbers could move

Engineering firmware, memory settings and power limits can all change before production silicon ships, as Fudzilla points out. Geekbench is also one workload mix, and its aggregate score says little about, say, a database or a virtualised cloud host.

Cache plays a part too. The 9996 has far more L3 per socket than the Zen 5 chips it’s compared against, and a 16-core Ryzen or EPYC 4545P has a very different cache layout. Some of the gain may come from architecture and some from the surroundings, and a single leaked run can’t separate them.

The Intel angle

The obvious rival is Intel’s Xeon 6+ “Clearwater Forest”, which launched at Computex 2026 with up to 288 Darkmont E-cores on the Intel 18A process. The flagship Xeon 6990E+ has 576MB of L3 and a 450W TDP, according to The Register’s spec table, and supports 12 channels of DDR5-8000. Intel’s own testing with Ericsson claims a single 288-core chip delivers 30% higher performance than a dual-socket Sierra Forest platform with the same total core count.

Intel wins on core count (288 against 256) and on shipping today. AMD’s leaked chip would have 512 threads against Intel’s 288, a bigger cache pool and, on this evidence, noticeably stronger cores. Those are leaked figures set against vendor-supplied ones, so don’t draw a verdict yet.

Price and availability

AMD hasn’t announced pricing or a release date for the EPYC 9996. For scale, the current flagship dense part, the 192-core EPYC 9965, launched at $14,813 per chip in 1,000-unit quantities (a US dollar list price; there’s no official EUR figure). Expect the 256-core part to land above that, though that’s an inference from AMD’s pricing history and not a leak.

Consumer chips are a separate story. Desktop Ryzen processors will use standard Zen 6 cores rather than Zen 6c, ShiftDelete reports, under the Olympic Ridge codename, and their cache, core count and clocks will differ from this server part.

Who should care

Cloud and hosting buyers are the direct audience: dense-core EPYCs are bought by the rack, where core count per socket and efficiency decide the purchase. A strong per-core gain on the same dense design helps AMD against Intel’s 288-core Xeon.

Everyone else gets a hint at the direction. Zen 5’s IPC uplift over Zen 4 was modest, and the clock gain was similarly small, so it was an unexciting upgrade for most desktop owners. Tom’s Hardware notes that Zen 6 may bring a meaningful IPC gain as well as a sizeable clock jump, if clock rumours hold. If both arrive, Ryzen 9000 owners will have a reason to look at the next generation.

Also this week

Intel’s Raptor Lake Next shows up in a driver and shipping logs. Tom’s Hardware found an internal Intel test driver labelled “RPL-NEXT” in Intel’s software database; the file is locked and tagged “testbuild”, so it probably wasn’t meant to be public. Club386 separately published customs data showing four Raptor Lake Next part numbers shipping from Vietnam since August 2026, including 99DLMV (3.5GHz, 30MB cache) and 99DLMT (2.1GHz, 33MB). Intel hasn’t confirmed specs or timing, though Gigabyte has indicated new LGA 1700 processors by early 2027 (via Wccftech).

Modders remove the half-rate tensor limit on RTX 20, 30 and 40 cards. Unofficial Linux kernel modules lift Nvidia’s limit that runs FP32-accumulate matrix math at half the FP16-accumulate rate on GeForce cards. BigGo reports BF16 throughput on an RTX 3090 Ti rising from 87.5 to 157.4 TFLOPS, with real-world gains of 14% to 36% in vLLM prefill and 47% to 59% in ComfyUI image generation; gaming and token decoding barely change. It’s Linux-only for now, Windows builds are experimental, and ai.rs found the trick doesn’t carry over to the RTX 5090.

What this means for you

If you’re planning a desktop upgrade, a leak from a 256-core server chip is no reason to hold off. Zen 6 desktop parts will have different core counts, cache and clocks, and AMD hasn’t given a date. Ryzen 9000 owners can safely sit tight until retail reviews exist.

If you buy servers or rent cloud capacity, note the 16.5% clock-matched figure and ignore the 24% headline until independent tests land. Ask vendors about Venice timing before signing multi-year hardware commitments, because Intel’s Xeon 6+ is the only 288-core option you can actually order.

If you’re on LGA 1700 with DDR4 and don’t want to pay for DDR5, keep an eye on Raptor Lake Next in early 2027, since it’s meant to drop into existing boards. Wait for official model names and prices before treating any shipping-log spec as real.

And if you run local models on an RTX 3090 or 4090 under Linux, the tensor unlock is worth trying on a spare machine first. It modifies GPU state before the driver loads, it’s unofficial, and the modules’ own documentation says a card already initialised by another driver is refused. Don’t try it on the box you depend on.