AMD's 256-core Zen 6 server chip ships, and early tests show big gains

Venice is a family of processors for data centers. These are the chips inside the big computers that run websites, databases and AI services.

A processor core is one worker that follows instructions. A thread is a lane of work that a core can juggle. More cores and threads mean more jobs at once.

AMD sells two flavors of the Zen 6 core:

• Zen 6: the full-speed version, up to 96 cores per chip, boosting up to 5 GHz.
• Zen 6c: a smaller version that runs slower but packs in far more cores, up to 256 per chip.

Both use the same core design. The "c" version is just trimmed to take up less space and use less power.

The top chip is the EPYC 9996. It has 256 Zen 6c cores, a 2.55 GHz base speed and a 4.1 GHz boost speed. It also carries 1,024 MB of cache, a small pool of very fast memory right on the chip.

Its default power rating is 600 watts, and it can be set as low as 400 watts. AMD lists the price at $14,904 per chip when bought in batches of 1,000.

Servers can hold two chips. A two-chip machine tested by reviewers had 512 cores and 1,024 threads in total.

In real life Picture a restaurant kitchen. Older chips were a kitchen with 192 cooks. The 9996 is the same kitchen with 256 cooks, a bigger pantry and faster doors between the rooms.

Zen 6 is built on TSMC's N2P process, the manufacturer's newest widely used way of making chips. At this scale, parts are only a few billionths of a meter wide. The process promises 18% more speed at the same power, or 36% less power at the same speed, compared with the one used for Zen 5's cores.

AMD says the design gets about 27% more work per tick of the clock than Zen 5, in enterprise, cloud and AI jobs. That number is not a promise for every program, and AMD measured it against a 16-core Zen 6c setup.

Other upgrades:

• Bigger building blocks: one block of cores can now hold 12 full cores or 32 smaller ones, up from 8 and 16.
• More cache: each small-core block carries 128 MB, four times the older design.
• Two I/O chips: the part that talks to memory and add-on cards is now split in two, which allows 16 memory channels per socket.
• Faster memory: up to 12,800 MT/s with special memory modules, and up to 1.6 TB per second of bandwidth per socket, AMD says.
• PCIe 6.0: the newest generation of the connection used by graphics cards and fast storage, at up to 64 Gbps per lane.

The catch: the extra memory channels come with a trade. The SP7 socket gets 96 PCIe 6.0 lanes, while the smaller SP8 socket offers 128.

Reviewers who got early hardware from AMD have now published results. In one industry test of throughput, a two-chip EPYC 9996 server scored 93% higher than a two-chip server with the previous 192-core flagship, the EPYC 9965.

Throughput means how much total work a machine finishes in a given time. It is the number that matters when a server handles thousands of requests at once.

AMD adds its own claims from its testing:

• Versus the last generation: about 2 times the integer compute, 2.4 times the floating-point compute and 2.7 times the memory bandwidth.
• Versus Intel: the 9996 delivers 1.8 to 3.13 times the speed of Intel's 128-core Xeon 6980P on molecular dynamics, materials modeling and weather forecasting.
• Web and databases: gains of 2.4 to 3.7 times across server-side Java, OpenSSL, MongoDB, Redis, NGINX and transaction processing.

AMD says the flagship delivers 2.24 times the platform-level performance of a Nvidia Vera system and 1.2 times its per-core performance.

Vendor claims deserve a grain of salt. Testers noted that Venice often leads, though Intel's next chip, Diamond Rapids, is due next year and should tighten the race.

Nvidia's Vera is a different kind of chip: an Arm-based processor with 88 cores, built for AI servers. AMD's closer rival to it is the EPYC 9686F, a 96-core Zen 6 part tuned for high clock speeds, up to 5 GHz.

Reviewers found the 9686F beats Vera in many tests, and shines in scientific computing. Vera still holds its own in a few spots, so the answer depends on the job.

AMD pitches Venice as a chip for agentic AI. An AI agent is a program that plans a task, calls tools and checks its work, instead of just answering one question. Each step runs code, searches data and moves files, and that work lands on regular processors, not graphics chips.

AMD added three features to make big server fleets easier to run:

• UPP (Unified Power Provisioning): shares the power budget across the whole machine. AMD says a rack that held 16 servers can hold 18 in the same power limit.
• UBPS (Utilization Based Power Scaling): adjusts power as the workload rises and falls.
• FAST (Frequency Adaptive Spectrum Technology): lets one chip run two speed tiers. In AMD's example, 32 cores of a 96-core chip run as a "premium" group up to 5 GHz, while the other 64 run as a slower "batch" group at about 2 GHz.

There is also MEXT, which AMD acquired this year. It lets a server use flash storage as extra memory, by moving rarely used pages out and guessing which ones come back. AMD says an app kept 95% of its speed with only 50% of the usual memory, and 80% with 25%. That claim is AMD's own, from internal use on 145 servers.

The SP7 chips are the first wave. The 31-chip lineup runs from about $700 to $14,904. Nine parts use the SP7 socket and 22 use the smaller SP8 socket.

SP8 machines, with 8 memory channels and more PCIe lanes per slot, arrive in the first half of 2027. A cache-heavy EPYC 9006X and a low-power Verano line follow in the second half of 2027. AMD's consumer Zen 6 chips, code-named Olympic Ridge, are also expected next year.

If you follow AMD, this lands after its purchase of World Labs, covered in our earlier story, and its push into local AI hardware, like the Gorgon Halo mini-PC.

• You can't buy it: Venice is for data centers. Your PC doesn't change.
• Cloud services could benefit: AMD says more servers fit in the same power budget, which is the kind of saving data centers look for.
• AI services get roomier: the busywork around chatbots runs on chips like these.
• Watch the independent tests: AMD's biggest numbers are its own. More reviews are coming.