Hardware/Comparison

AMD Zen 6 vs Zen 5: IPC, Cores and Cache Compared

AMD has published real Zen 6 architecture numbers for the first time. Here is exactly what changed versus Zen 5 — and what is still EPYC-only, not Ryzen-confirmed.

An AMD EPYC processor seated in its server socket, representing the Zen 6 architecture that powers AMD's EPYC "Venice" CPUs
An AMD EPYC processor in its server socket. Image: AMD.

AMD has now put real numbers behind Zen 6, and they land on the server side first: a 27% geomean IPC uplift in AMD's own published testing, compute chiplets (CCDs) that grow from 8 cores to 12, L3 cache per CCD rising from 32MB to 48MB at the same 4MB-per-core ratio, and a move to TSMC's 2nm-class N2 process. All of that comes from AMD's EPYC™ 9006 "Venice" disclosures, not from a Ryzen desktop briefing — AMD has not yet published Zen 6 core counts, cache sizes, clocks or IPC figures for its consumer chips, so anything you read about 24-core "Olympic Ridge" Ryzen parts is still outside analysts' rumors, not AMD's own word.

Zen 6 vs Zen 5, what AMD has actually confirmed:

  • CPU process: TSMC 2nm-class N2, up from 4nm on Zen 5 — AMD's first CPU architecture on N2.
  • IPC: AMD cites a 27% overall geomean gain for Zen 6c vs. its Zen 5c predecessor at a matched, fixed clock — with a 31% gain on AI workloads and 20% on enterprise/cloud workloads.
  • Compute chiplet (CCD) cores: up to 12 on standard Zen 6, up from 8 on Zen 5 — a 50% increase.
  • L3 cache per CCD: up to 48MB on Zen 6, up from 32MB on Zen 5, holding the same 4MB-per-core ratio.
  • All of the above is from EPYC "Venice" server disclosures. Desktop Ryzen Zen 6 specifics are still unconfirmed by AMD.

What AMD actually confirmed about Zen 6

Unlike the usual run of Zen 6 leaks, the figures in this piece come from material AMD itself has published. AMD's own "Zen" Core Architecture page now carries a Zen 6 entry in its generation-by-generation IPC table, and AMD's AMD EPYC Server CPU Architecture and Performance Overview whitepaper lays out the chiplet design, core-count ceilings and vector-compute story behind the 6th Gen EPYC 9006 "Venice" family that Zen 6 and its dense sibling Zen 6c power. Both documents are dated to AMD's November 2026 disclosure window, alongside the company's own newsroom confirmation that "Venice" has begun production on TSMC's 2nm process.

Here's the catch worth keeping in view throughout: every one of those documents is written around EPYC, AMD's server CPU line. Zen 6 and Zen 6c are the same underlying core designs that will eventually anchor Ryzen desktop chips, but AMD has deliberately disclosed the architecture through its server and data-center materials first. That's a very different posture from, say, Zen 5's 2024 unveiling, which arrived through Ryzen 9000 marketing before EPYC "Turin" followed. We'll come back to what that means for desktop buyers after walking through the numbers.

Zen 6 vs. Zen 5: the architecture scorecard

The table below pulls together what AMD has published for Zen 5 (the Ryzen 9000 and EPYC "Turin" generation) against what it has now disclosed for Zen 6 (EPYC "Venice"). Where a figure is AMD's own confirmed number, it's marked; where it's a reasonable read of AMD's own charts cross-checked against independent technical reporting, that's flagged too.

SpecZen 5 (Ryzen 9000 / EPYC 9005)Zen 6 (EPYC 9006 "Venice")
CPU process node4nm / 3nmTSMC 2nm-class N2 (AMD-confirmed)
Standard CCD core count8 coresUp to 12 cores (+50%, AMD chart + press corroboration)
Dense "c" CCD core count16 cores (Zen 5c)Up to 32 cores (derived from AMD's 256-core max ÷ 8 CCDs, press-corroborated)
L3 cache per standard CCD32MB48MB (AMD whitepaper diagram)
L3 cache per dense "c" CCDUp to 32MB (Zen 5c)128MB (Zen 6c, AMD whitepaper diagram)
L3-per-core ratio4MB/core4MB/core (unchanged)
Max EPYC cores per socket192 (Zen 5c)96 (Zen 6) / 256 (Zen 6c), AMD-confirmed
IPC vs. prior generation~16% (Ryzen, AMD-confirmed) / ~17–37% (EPYC, AMD-confirmed)~20% enterprise / ~31% AI geomean (EPYC, AMD-confirmed); ~27% overall in AMD's matched 16-core test
AVX-512 FP16 / VNNI-INT8FP16 conversion only; 2 VNNI-INT8 pipelinesFull AVX-512 FP16 math; 4 VNNI-INT8 pipelines (per AMD's Zen 6 technical briefing, as reported)
Ryzen desktop specificsFully shipped and documentedNot yet published by AMD

The 27% IPC figure: what it measures, and what it doesn't

The headline "27% faster per clock" line that's been circulating since AMD's Zen 6 technical disclosures in Las Vegas in early November 2026 has a very specific, footnoted origin, and it's worth reading before repeating it. AMD's own Zen architecture comparison page attaches footnote "9xx6-062A," dated November 5, 2026, which lays out the exact methodology: a geomean IPC comparison, at a fixed 3GHz, between an EPYC 9006-series "9996" part (Zen 6c, 16 cores active on a single CCD) and an EPYC 9005-series "9B45" part (Zen 5c, 16 cores active across two 8-core CCDs). Across 27 workloads in seven categories, AMD's measured results were: server kernels 1.39x, AI 1.31x, database 1.28x, HPC 1.27x, cloud-native 1.25x, web serving 1.22x, and industry-standard benchmarks (SPECrate, SPECjbb, SPECpower) 1.20x — for an overall geometric mean of 1.27x, AMD's own "27% uplift."

Two things fall out of that footnote that matter for anyone comparing Zen 6 to Zen 5. First, this is a Zen 6c vs. Zen 5c comparison, not standard Zen 6 against standard Zen 5 — AMD's own summary text on the same page separately describes "Zen 6c" delivering "approximately 31% higher IPC across a geomean of AI workloads and 20% higher IPC across a geomean of industry-standard enterprise and cloud benchmarks," which AMD presents as the headline Zen 6 generation-over-generation figure in its own IPC table. Second, it's a fixed-frequency, same-core-count EPYC comparison designed to isolate microarchitecture from clock speed and core count — useful for judging the redesign itself, but not a number AMD has extended to gaming, content-creation or any other desktop-style workload.

Guru3d and Tom's Hardware, both of whom attended AMD's technical session, reported that the uplift comes from a genuinely wider core: AMD widened Zen 5's front end, execution ports and load/store bandwidth rather than making incremental tweaks, with a bigger reorder buffer, wider dispatch and larger FP/INT schedulers cited as the main contributors. AMD representatives reportedly cautioned at the same session that Ryzen results "may differ" from the EPYC-measured figures, since desktop workloads, clock targets and core configurations are all different from the server parts AMD has tested so far.

Bigger chiplets: 12-core CCDs, up 50% from Zen 5's eight

AMD's own EPYC whitepaper diagrams the "Venice" package as compute chiplets (CCDs) running "Zen 6" or "Zen 6c" cores surrounding dedicated I/O die logic, the same chiplet-plus-IOD approach AMD has used since the original "Zen" and explains in its architecture materials as a way to let core and I/O silicon use different process nodes and validation schedules. AMD confirms Zen 6's EPYC ceiling at up to 96 cores per socket and Zen 6c's at up to 256 — and AMD's own chiplet diagram shows eight CCDs doing that work on each design. Divide it out and you get 12 cores per standard Zen 6 CCD (96 ÷ 8) and 32 cores per Zen 6c CCD (256 ÷ 8), both of which line up with what independent outlets, including Tom's Hardware, Guru3d and Club386, reported AMD stating directly in its Zen 6 architecture presentation: a jump from Zen 5's 8-core CCD to a 12-core Zen 6 CCD, a 50% increase in per-chiplet core density.

That density gain is also the most direct technical link to the desktop-chip speculation that followed AMD's presentation. Two 12-core CCDs would, on paper, put a future flagship "Olympic Ridge" Ryzen desktop part at 24 cores and 48 threads — but that specific product configuration is industry-press arithmetic, not an AMD announcement, and it's exactly the kind of number this piece is flagging as unconfirmed rather than repeating as fact.

L3 cache: 48MB per CCD, same ratio as Zen 5

AMD's chiplet diagram in its own EPYC 9006 whitepaper labels the standard Zen 6 CCD's L3 cache at 48MB, and the dense Zen 6c CCD's at 128MB — both before AMD's separate 3D V-Cache stacking technology is added on top (AMD's materials cite EPYC configurations reaching up to 1,152MB of total L3 with 3D V-Cache in place). Measured against Zen 5's well-documented 8-core, 32MB CCD, that's a straightforward consequence of more cores sharing the same cache architecture at an unchanged 4MB-per-core allocation: 12 cores x 4MB = 48MB for Zen 6, and 32 cores x 4MB = 128MB for Zen 6c. AMD isn't claiming a cache-efficiency breakthrough here so much as scaling a cache design it already trusts across a larger chiplet — which, if the ratio really does carry over unchanged to Ryzen desktop parts down the line, would be good news for cache-sensitive workloads like gaming, the same category AMD has leaned on heavily in the current flagship gaming matchup between its X3D-equipped chips.

A wider execution engine: schedulers, VNNI-INT8 and AVX-512 FP16

Beyond the headline core and cache numbers, AMD's Zen 6 technical session reportedly detailed several execution-engine changes aimed squarely at AI and vector-heavy workloads, as covered by Tom's Hardware and Club386. Those outlets describe an 8-wide dispatch core with larger floating-point and integer schedulers, a bigger return-address stack (growing from 52 to 64 entries per Guru3d's account), and improved branch prediction accuracy layered on top of the wider front end. On the vector side, Zen 6 reportedly doubles AMD's VNNI-INT8 pipeline count from two to four and adds full native AVX-512 FP16 math — Zen 5 could only convert FP16 data before computing on it in FP32, while Zen 6's pipelines are said to compute FP16 natively.

AMD's own whitepaper doesn't spell out those exact pipeline counts, but it does confirm the broader direction: a dedicated whitepaper section on "AVX-512, Vector Compute and Future ACE Support" describing AVX-512 as central to AI, analytics, scientific-computing and compression workloads on EPYC 9006, alongside AMD's participation (with Intel, Microsoft, Google Cloud, Meta and others) in the x86 Ecosystem Advisory Group that's standardizing a future AVX10 vector extension and a unified "ACE" matrix-multiplication approach across the industry — context that makes clear why AMD is investing so heavily in wider vector pipelines right now.

TSMC 2nm: Zen 6's new foundation

The process-node jump is the one piece of this story with the least ambiguity. AMD's official EPYC comparison table lists Zen 6/Zen 6c's process as "2nm," up from "4/3nm" for Zen 5, and AMD's own newsroom release states plainly that "Venice" is "the first HPC product in the industry to achieve production ramp on TSMC advanced 2nm technology," quoting CEO Lisa Su on the milestone and TSMC CEO C.C. Wei confirming the partnership from TSMC's side. TSMC's own technology pages describe the underlying N2 node as its first to use gate-all-around nanosheet transistors, a structural change from the FinFETs used on N3/N4, which TSMC says delivers a "full-node" jump in performance, power and transistor density versus its previous processes. Wikipedia's entry on the Zen microarchitecture family has the fuller generational lineage if you want to see how each node transition lines up against past Zen releases.

It's worth separating two different claims here: that Zen 6 is built on N2 (AMD-confirmed, unambiguous), and that N2 itself delivers any particular performance or efficiency percentage (a TSMC process claim, not something AMD has broken out as "this many percentage points of Zen 6's IPC gain came from the node versus the microarchitecture"). The 27%, 31% and 20% IPC figures above are AMD's architecture-level claims; the node is one input into them, not a separately quantified one.

Why none of this is a Ryzen spec sheet yet

Every figure in this piece traces back to EPYC "Venice" material: AMD's own newsroom, its EPYC whitepaper, and the EPYC column of its Zen architecture comparison table. The Ryzen column of that same AMD table is the clearest evidence of the gap — it still tops out at "AMD Ryzen 9000," running "Zen 5," with a ~16% IPC figure, and simply has no Zen 6 entry yet. AMD has not published a Ryzen-specific core count, L3 cache size, clock speed or IPC number for Zen 6 desktop chips as of this writing.

That gap lines up with the timeline this desk covered separately: AMD's desktop Zen 6 Ryzen chips have reportedly slipped to 2027, a story with its own set of sourcing caveats that we walked through in our look at the rumored desktop delay. Combined with AMD disclosing the architecture through EPYC first, the pattern suggests Ryzen-specific marketing, core counts and clock targets are still some months out — and that the 24-core "Olympic Ridge" flagship figure, the specific Ryzen clock speeds, and any gaming-IPC percentage you see attributed to Zen 6 right now should all be read as informed estimates built on the EPYC disclosures in this piece, not as AMD's own desktop roadmap.

What's next for Zen 6

The near-term calendar is clearer on the server side than the desktop one. AMD's EPYC 9006 "Venice" SP7 parts are already in production on TSMC's 2nm process and shipping to server makers, with the enterprise-focused SP8 variant and the LPDDR-equipped "Verano" follow-on expected through 2027 per AMD's own roadmap language, and AMD's newsroom has already begun detailing how it expects EPYC 9006 to serve agentic AI infrastructure workloads in production. On the desktop side, the next milestones to watch are AMD actually populating the Ryzen column of its own Zen architecture comparison table with a Zen 6 entry, and any CES 2027 or subsequent AMD keynote that puts real Ryzen core counts, clocks and pricing next to the EPYC numbers detailed here. Until AMD does that directly, treat every "Ryzen Zen 6" spec you see — this piece's own chiplet math included — as derived from server disclosures rather than confirmed consumer hardware.

Frequently asked questions

Has AMD confirmed Zen 6 IPC gains for desktop Ryzen chips?

No. AMD's own Zen architecture comparison page lists IPC figures through "AMD Ryzen 9000" (Zen 5) with no Zen 6 row yet. Every core-count, cache and IPC figure AMD has published so far comes from its EPYC "Venice" server disclosures, not from Ryzen.

What is AMD's official Zen 6 IPC uplift claim?

AMD cites roughly 31% higher IPC on a geomean of AI workloads and 20% higher IPC on enterprise/cloud workloads for Zen 6c versus Zen 5c. A separate, footnoted AMD test comparing matched 16-core EPYC configurations at a fixed 3GHz measured an overall geomean uplift of 27%.

How many cores does a Zen 6 compute chiplet (CCD) have?

Dividing AMD's own published EPYC core-count ceilings by the chiplet count in AMD's chiplet diagrams works out to 12 cores per standard Zen 6 CCD, up from 8 on Zen 5, which matches what independent outlets reported AMD stating directly at its Zen 6 technical session.

How much L3 cache does Zen 6 add over Zen 5?

AMD's EPYC whitepaper diagrams show 48MB of L3 on a standard Zen 6 compute chiplet, up from 32MB on Zen 5, keeping the same 4MB-per-core ratio. The denser Zen 6c chiplet carries 128MB.

What manufacturing process is Zen 6 built on?

TSMC's 2nm-class N2 node — AMD's first CPU architecture on N2, and what AMD's own newsroom calls the first HPC product in the industry to reach production ramp on that process.

When will Zen 6 desktop Ryzen chips launch?

AMD hasn't given a desktop date yet. Reporting on AMD's roadmap points to a 2027 arrival for Ryzen Zen 6 chips, later than earlier expectations — see our separate coverage of the rumored delay for the sourcing behind that timeline.

Sources

More on AMD Zen 6 →AMDZen 6Zen 5CPU architectureEPYC VeniceTSMC 2nm
Mara Lindqvist
Written byMara Lindqvist

Mara Lindqvist edits the hardware desk. She covers graphics cards, processors, memory and storage, the foundries and chip designers behind them, and what the numbers on a spec sheet mean for people choosing a PC. Specifications in her stories come from manufacturer spec pages and datasheets.

More from Hardware

See all