AMD's Zen 6 Architect on Why Venice Cut a Multiplier and Added Vector Loads
Hardware / analysis
AMD's Zen 6 Architect on Why Venice Cut a Multiplier and Added Vector Loads
Mike Clark's Chips and Cheese interview and The Next Platform's launch tables show a dense 32-core chiplet promoted to the performance core.

AMD's sixth-generation EPYC 9006 "Venice" family launched Nov. 10 with 31 processors from 8 to 256 cores, priced from $700 to $14,904 per chip in 1,000-unit trays, according to Timothy Prickett Morgan at The Next Platform. A day later Mike Clark, who describes himself as "the original architect of Zen", told Chips and Cheese what changed inside the core, in a transcript published Nov. 11. The answer is a set of trade-offs rather than a wider machine.

An early Ryzen, the first Zen part, shown for reference. Venice is a server chip.
The dense core is now the performance core
The headline architectural claim is Clark's: "our dense technology is now our top-of-the-line, no-questions-asked performance core." Zen 6c, the dense variant, sits in a 32-core chiplet with 128 MB of L3, or 4 MB per core, while classic Zen 6 is the high-frequency part in a 12-core chiplet with 48 MB. Both chiplets are TSMC 2 nm; the I/O die is 6 nm.
AMD wants the two cores to behave identically so software needs no tuning for one or the other. Zen 6c compensates for its lower clock target with larger queues, including the floating-point scheduler, which Clark says buys IPC under SMT. Chips and Cheese's George Cozma says Zen 6c matches the fastest Zen 5 server CPU in SPECint 2026 despite the lower clock. That is the interviewer's statement, not an AMD figure.
What was cut: the third multiplier
Integer multiply units drop from three on Zen 5 to two. Clark says the third multiplier added little on Zen 5 and AMD cut it to save area, returning to more distributed schedulers that feed the ALUs directly. The sentence is worth the attention because spending area on a unit that rarely helped is the kind of decision a spec sheet never shows.
Vector loads go the other way. Zen 5 handled two per cycle; Zen 6 handles four at 128 or 256 bits, reusing wiring that already existed for two 512-bit loads. Two more VNNI units were added late, to meet inference demand. Clark's own limit on the claim: "It's not going to beat a GPU, and that's not the point."
Cozma measured the L2 delivering 64 bytes per cycle on reads. Clark attributes Zen 5's shortfall to over-prefetching, where prefetches that hit in L2 consumed bandwidth demand requests needed.
Platform numbers: SP7 and SP8
The Next Platform's platform comparison is the part to price against. Per-socket figures:
| Attribute | SP7 | SP8 |
|---|---|---|
| Status | Shipping now | Expected first half of 2027 |
| DDR5 channels | 16 | 8 |
| PCIe lanes | 96 PCIe 6.0 plus 8 PCIe 4.0 | 128 PCIe 6.0 plus 8 PCIe 4.0 |
SP8 memory bandwidth is half of SP7's. On SP7, bandwidth is up to 1.6 TB/s with MRDIMMs at 12.8 GT/s and 1 TB/s with RDIMMs at 8 GT/s.
- MRDIMM at 12.8 GT/s1.6 TB/s
- RDIMM at 8 GT/s1 TB/s
Source: The Next Platform, Nov. 10, 2026
The denominator that matters is bandwidth per core. If the top SP7 part carries 256 cores, 1 TB/s of RDIMM bandwidth works out to about 3.9 GB/s per core and 1.6 TB/s of MRDIMM bandwidth to about 6.3 GB/s (our arithmetic, not AMD's). That is why the MRDIMM figure matters to anyone buying dense parts for bandwidth-hungry work.
Branch prediction and front end
Two smaller changes round out the interview. Zen 5 introduced dual decode, fetching two basic blocks at once; Zen 6 optimises how a single thread uses both decode clusters when a workload draws on the instruction cache and the op cache together, aiming at less wasted power. Misprediction recovery takes fewer cycles than on Zen 5, and prediction accuracy is at least as good, Clark says, with the TAGE tables laid out more efficiently and no bigger.
Clark also said AI tools helped with some of the optimisation work. The interview does not give a count of how many design decisions that covers.
The vendor numbers to hold at arm's length
AMD claims about a 20 percent per-core integer uplift over Zen 5 at the same clock, per The Next Platform. It also claims the 256-core EPYC 9996 delivers about 190 times the performance of a 2008 Opteron 2393 SE at 14.9 times lower cost per unit of performance. Both are vendor-supplied, and the 2008 baseline flatters any modern part. Chips and Cheese reports that Clark withheld some L3 details, and the interview gives no clocks or process-node figures.
Clark's remark on the late VNNI addition shows where AMD sees the CPU: as the host that runs inference it cannot hand off. Nvidia, whose chief executive just changed (our report), sells the accelerators that sit beside these hosts. Start-ups chasing that workload with new silicon, such as Fractile, are betting the host will not be enough.
What would change this read is independent SPEC and memory-latency runs on the 32-core chiplet. Clark says it has slightly higher L3 latency than the 12-core one and that bigger queues hide it. Third-party numbers will show whether they do. Until then, treat the efficiency story as AMD's.
Sources
More in Hardware
- 01Anthropic's 87% Gas Power Figure Is a Model of 9.6 GW Drawn From a 15 GW CommitmentThe Better Data Center Project's estimate rests on 5.9 to 9.6 GW with named sites, 2.5 GW of on-site gas and a $3.4 billion generator purchase reported by the Financial Times.
- 02Serve Robotics Ran 917 Active Robots and Lost $7.6 Million GrossThe sidewalk-delivery company says it has deployed more than 2,000 machines, and its quarterly release sets $10.4 million of cost of revenues against $2.8 million of sales.
- 03Rockwell Automation Agrees to Buy Austria's AGILOX, Terms UndisclosedThe deal pairs AGILOX's forklift robots with Rockwell's OTTO flatbeds, and neither company says what Rockwell is paying.
- 04Italy's Tesla Approval Roughly Doubles the EU Population Covered, to Close to 30%, Before December's VoteItaly is the 12th European country to recognise the Dutch RDW approval of Tesla's supervised driving system. The share of EU residents who can legally use it hinges on one number: 59 million.