Test report DSG-2796 · Rev C · tested November 17, 2026

Processors & AcceleratorsDevice under test

AMD's Zen 6 Epyc 9006 hits 2,600 GB/s memory bandwidth

AMD's Zen 6 Epyc 9006 launches with 256 cores and 2.5x memory bandwidth. First tests: dual 9686F hits 2,600 GB/s in STREAM, doubling Intel's Xeon 6980P duo.

Read
5 min
Words
950
Node
65nm
Operator
Elena Vasquez

Spec summary

  1. Dual Epyc 9686F reached ~2,600 GB/s STREAM bandwidth vs 1,200 GB/s for two Intel Xeon 6980P.
  2. Epyc 9006 scales to 256 cores / 512 threads; 16-channel DDR5 MRDIMM at 12,800 MT/s delivers 2.5x predecessor bandwidth.
  3. AMD claims up to 80 percent more tokens per second for agentic AI versus Epyc 9005.
  4. SP8 socket with 22 models follows H1 2027; Zen 7 'Florence' planned for 2028, Zen 8 'Ravenna' for 2030.
  5. Entry 8-core Epyc 9016 costs $700 wholesale; flagship 256-core 9996 lists at $14,904.
AMD Zen 6 soll agentische KI beflügeln: Erste Leistungswerte
Fig. AAMD Zen 6 soll agentische KI beflügeln: Erste Leistungswerte — Lalmch / pixabay

A dual-socket Epyc 9686F server reached almost 2,600 GB/s in the STREAM Triad benchmark — more than double the 1,200 GB/s of two Intel Xeon 6980P with 12-channel DDR5-8800 MRDIMM. That's the headline number from AMD's Zen 6 launch, and c't's remote measurements suggest the architecture delivers beyond raw bandwidth too.

The Epyc 9006 family is rolling out in stages. Nine regular SP7 CPUs plus the Helios-exclusive 9G76 and six embedded models are first; AMD says it's already shipping, with first systems due this year. The compact SP8 socket follows in H1 2027 with 22 models, targeting general-purpose and cloud servers. After mid-2027 come Venice-X with 3D stacked cache for HPC and Verano LP with 192 cores and LPDDR memory for the next AI server generation. Zen 6 will also land in desktop and mobile chips — rumors point to Ryzen 10000, but AMD isn't talking.

What do the first benchmarks show?

The test server ran two Epyc 9686F at 96 cores each, up to 5.3 GHz, with 2 TB of RAM in 32 DDR5 MRDIMMs at 12,800 MT/s. For comparison: the previous Epyc 9005 topped out at 908 GB/s with 12-channel DDR5-6000.

The bandwidth gain translated directly into throughput. In 7-Zip compression, the 9686F pair became the first CPU duo in the test lab to break 2 GB/s of compression speed.

Kernel compilation with kcbench 0.9.16 told a similar story. The 9686F pair beat two Epyc 9755 — chips with over a third more cores — even with Clang 21.1.8, which scales better with core count than GCC 16.0.1. Blender 5.2.3 LTS rendering in the classroom, benchmark and barbershop scenes put the Zen 6 tandem up to 20 percent ahead of the on-paper-faster Epyc 9755.

Single-threaded, one Zen 6 core at 5.3 GHz finished the BMW27 scene in 286 seconds. The 5.0 GHz Zen 5 core in the Epyc 9575F needed 349 seconds — roughly 22 percent longer. The higher TDP of the 9686F plays no role here.

How much faster is Zen 6 really?

AMD claims a 27 percent average uplift for a 16-core Zen 6c versus Zen 5. But the comparison pits a two-CCD8 Zen 5 setup with I/O-die detours against a Zen 6c CCD32 where everything stays on-die. Footnotes were missing from the PDF, so differing L3 sizes and memory speeds may skew the picture. The slide says "IPC uplift," but IPC claims have rarely been literal in cross-generation comparisons.

The architecture itself evolved from Zen 5 rather than replacing it:

  • Op-cache grew from 6,000 to 6,750 entries, with dispatch bandwidth up 50 percent to 18 micro-ops per cycle.
  • The integer scheduler keeps 144 entries but redistributes them: 48 fixed ALU slots plus 96 shared with the AGUs (Zen 5: 88 ALU, 56 AGU fixed).
  • FP scheduler queue grew from 114 to 180 entries on Zen 6c, likely 138 on Zen 6.
  • Load and store queues grew 25 and 33 percent; the register file grew 47 percent to 352 entries; the reorder buffer now holds 640 entries.

Zen 6 and Zen 6c share 4 MB of L3 per core. Zen 6c packs 32 cores plus 128 MB L3 into one CCD32; high-clock Zen 6 cores share a CCD12. I/O dies are now split — two per package — built on TSMC's 6 nm process, while the CCDs use the 2 nm node. Chief architect Mike Clark admitted clock-speed trade-offs: "We cut some structures down […] to get to that frequency, but we believe, they're very minimal." One visible difference: FP scheduler counts of 180 (3×60) on Zen 6c versus at least 138 (3×46+) on Zen 6. Software optimization, Clark said, stays the same.

What does the platform change?

SP7 measures 76×80 mm and supplies up to 600 W, with 16 DDR5 channels at 1 DPC — DDR5-8000 or MRDIMM at 12,800 MT/s. That's 1.6 TB/s per socket, 2.5x the Epyc 9005 rate. AMD promises up to 80 percent more tokens per second for agentic AI workloads versus the previous generation.

Clocks reach 5.3 GHz on Zen 6 (6 percent over the 5.0 GHz predecessors) and 4.1 GHz on Zen 6c (about 11 percent). PCIe 6.0 doubles per-lane bandwidth: SP7 offers 96 lanes single-socket, 128 dual; SP8 tops out at 128 and 192 respectively. CXL 3.1 is supported.

Max core count is 256 — 512 threads with SMT, active on all announced models. At the bottom, an 8-core part costs $700 wholesale (around €625). The 256-core 9996 runs $14,904. SP8, arriving 2027, is capped at 128 cores and 400 W with eight channels, though 2 DPC matches SP7's memory capacity at lower speed. Embedded SP7 variants get seven years of support, with the next CPU generation fitting the same boards.

Model numbering: the first digit marks the 9000 family, the second encodes core count (9: 192–256, down to 0: 8), the third relative performance, the last the generation. Suffixes add traits — P for single-socket, F for high frequency, N for embedded/edge, X for 3D V-Cache (H2 2027).

What's next on the roadmap?

Zen 7, codename "Florence," is planned for 2028 with first confirmed support for the AMD-Intel standardized ACE matrix extension and "next gen memory" via MRDIMM and LPDDR — DDR6 isn't named outright. Zen 8 "Ravenna" is in development for 2030, with no details disclosed.

via heise.de/rss/heise-top-atom.xml (Source)

Filed under

  • amd-epyc
  • zen-6
  • server-processors
  • memory-bandwidth
  • ddr5-mrdimm
Share this article:

More from Elena Vasquez

Elena Vasquez

Show full bio

Senior reporter covering people, launches and day-to-day practice in hardware layer of artificial intelligence at Die Signal.

675 articles

Same lot · LOT-362D

« Previous articleNext article »