AMD lifted the review embargo on its Epyc 9006 server processors on 10 November, and the headline is not the core count. It is the memory pipe. In a dual-socket configuration, two 96-core Epyc 9686F parts running 2 TB of DDR5 MRDIMM at 12,800 MT/s pushed the STREAM Triad benchmark to almost 2,600 GB/s — against 1,200 GB/s for a pair of Intel Xeon 6980P and about 908 GB/s for the previous Epyc 9005 generation at DDR5-6000.

That is not a rounding-error improvement. It is roughly a doubling of the number Intel's current top bin can reach, and it is the figure AMD is leaning on hardest, because it is the figure that decides how many tokens per second a rack of inference servers can produce when the model no longer fits in cache.

The architecture story behind it is a socket change and a process change at the same time. Epyc 9006 needs new boards: SP7 for the big version, with up to 256 cores, 600 watts and 16 memory channels at one DIMM per channel; SP8 for the compact version, capped at 128 cores and 400 watts with eight channels, arriving in the first half of 2027. Compute chiplets move to TSMC's 2 nm process; the I/O dies stay at 6 nm, and there are now two of them per package.

The core split follows the pattern set by Zen 5. The 96- and 64-core high-frequency parts (the F suffix) use full Zen 6 cores, twelve to a chiplet, topping out at 5.3 GHz. Everything denser uses Zen 6c, up to 32 cores with 128 MB of L3 on a single chiplet, at up to 4.1 GHz. AMD claims a 27% IPC uplift for Zen 6c over Zen 5 — a figure worth reading with the footnotes attached, since it compares a two-chiplet Zen 5 configuration against a single 32-core Zen 6c chiplet that avoids the I/O die entirely.

The first independent numbers are less arguable. A dual 9686F system became the first CPU pair in c't's lab to break 2 GB/s in the 7-Zip compression benchmark. The same pair beat two Epyc 9755 chips — which have a third more cores — when compiling the Linux kernel, even with Clang, the compiler that scales better with core count. In Blender rendering it led the 9755 by up to 20%. In a single-threaded scene, one 5.3 GHz Zen 6 core finished in 286 seconds where a 5.0 GHz Zen 5 core took 349 — about 22% faster.

AMD's own framing is that agentic AI is the workload this generation was tuned for, and it claims up to 80% more tokens per second than Epyc 9005. Treat vendor figures as vendor figures; the memory bandwidth is the part that is hard to argue with, because it is a property of the platform rather than of a benchmark harness.

Prices run from $700 for an eight-core Epyc 9016 to $14,904 for the 256-core Epyc 9996. Systems are shipping now, SP8 follows in the first half of 2027, and the 3D V-Cache Venice-X parts and the 192-core LPDDR Verano LP are due in the second half. Zen 7, codenamed Florence, is pencilled in for 2028.