AMD 6th Gen EPYC 9006 Venice CPUs reportedly offer up to 96 Zen 6 or 256 Zen 6c cores

(Image credit: AMD)

Specifications for AMD's next-generation Zen 6-based server products have surfaced at Baidu, via HXL at X. Venice, the successor to AMD's current Turin server offerings, will reportedly pack as many as 256 cores based on the Zen 6c design, suggested with 1GB of total L3 cache, and that's without any additional 3D cache chiplets. We must tread with caution and not treat these details as definitive, since these products are likely a year or more away from launch.

AMD's current-generation Zen 5-based server EPYC 9005 CPUs, codenamed Turin, slot into the SP5 socket, similar to EPYC 9004 (Genoa), which used Zen 4 cores, adopting a design similar to Arm's big.LITTLE. AMD also launched these CPUs with denser Zen 4c/5c cores, with Turin Dense reaching up to 192 cores/384 threads. Similarly, AMD also offers the cost-effective SP6 socket, which currently supports AMD's EPYC 8004 (Siena) offerings with up to 64 Zen 4c cores, also rumored to be compatible with the speculated Zen 5-based EPYC 8005 (Sorano) family.

AMD is rumored to introduce new SP7 and SP8 sockets for its next-generation server offerings, succeeding SP5 and SP6, respectively. The larger SP7 platform is expected to accommodate EPYC 9006 (Venice) CPUs, which reportedly feature as many as 256 Zen 6c cores, divided across eight CCDs. This lands us at 32-core Zen 6c CCDs, which are also purported to feature 128MB of L3 cache, for 1GB of total L3 cache (128MB x 8) for a 256-core chip.

While the top core count for standard Zen 6-based Venice CPUs on SP7 remains unknown, the leak suggests a significant update in the CCD design. Each Zen 6 CCD allegedly carries 12 cores and 48MB of L3 cache, a notable increase from the eight cores and 32MB seen since Zen 2.

The lower-cost SP8 platform will allegedly support four 32-core Zen 6c CCDs for 128 cores and 512MB of total L3 cache. That's a solid increase over Siena and likely even Sorano. Standard Zen 6 options on SP8 are projected to offer 96 cores (eight 12-core chiplets) and 384MB of L3 cache. This significant improvement in core density is likely a direct consequence of adopting TSMC's 2nm (N2) process node. Although specifics have not been detailed, we can expect similar improvements in I/O capabilities, including more memory channels and increased PCIe lane counts.

The rationale behind pursuing more cores and larger caches is evident in light of upcoming Diamond Rapids and Clearwater Forest, which are Xeon processors, anticipated to be some of Intel's most sophisticated and advanced designs in recent history. Likewise, AMD is also expected to shift to advanced packaging options from TSMC with Zen 6, potentially employing silicon interposers (TSMC's CoWoS-S) or silicon bridges (TSMC's InFO_LSI or CoWoS-L) to link its CCDs and IODs. Following the typical two-year cadence between generations, we can expect the first Zen 6-based products to arrive by H2 2026.

Follow Tom's Hardware on Google News to get our up-to-date news, analysis, and reviews in your feeds. Make sure to click the Follow button.

Hassam Nasir
Contributing Writer

Hassam Nasir is a die-hard hardware enthusiast with years of experience as a tech editor and writer, focusing on detailed CPU comparisons and general hardware news. When he’s not working, you’ll find him bending tubes for his ever-evolving custom water-loop gaming rig or benchmarking the latest CPUs and GPUs just for fun.

  • ejolson
    The article states "adopting a design similar to Arm's big.LITTLE"

    My understanding is that big.LITTLE is about putting different kinds of cores in the same packages to create a heterogeneous processor that can run light tasks using low power in mobile devices. This is not what's going on in the datacenter with Zen cores. Each core in an Epyc processor is the same.
    Reply
  • jeremyj_83
    Unless the 128c Zen 5 isn't selling well I don't see why AMD would drop for 128c > 96c for the halo Zen 6. I hope they go 16 channel RAM for SP7 and 8 channel for SP8. An increase in total PCIe lanes would also be nice. Getting at least a 160 on both single and dual socket would be good. Ideally 192 or more though as ES.1 form factor is gaining market share and allows for denser storage and a bunch of high speed networking.
    Reply
  • usertests
    jeremyj_83 said:
    Unless the 128c Zen 5 isn't selling well I don't see why AMD would drop for 128c > 96c for the halo Zen 6.
    I don't think it's dropped. It's just the limit for the smaller socket (SP8). Zen 6 core count for the big socket wasn't specified.

    Videocardz lists 128-core Zen 6 for SP7. Where they are getting that info, or if it's a guess, I'm not sure. They also have 16-channel for SP7, 4/8-channel for SP8.

    SP6 is the current cheap socket, supporting up to 64 Zen 4c cores. So 96-core Zen 6 would be a big upgrade (the article says there will be a Zen 5 Sorano for SP6 though). I think AMD realized that some customers want the faster cores on the cheaper socket. So they are allowing the choice of one or the other on either socket.
    Reply
  • usertests
    ejolson said:
    My understanding is that big.LITTLE is about putting different kinds of cores in the same packages to create a heterogeneous processor that can run light tasks using low power in mobile devices. This is not what's going on in the datacenter with Zen cores. Each core in an Epyc processor is the same.
    If AMD wanted to mix chiplets in a "hybrid" Epyc processor, or a customer wanted that, they could probably do it. Having 12 big and 224 small cores might make sense for some servers. But there's no evidence that AMD is planning to make that.

    Even in mainstream mobile APUs, Strix Point 4+8 looks like a stopgap because they weren't ready to switch to chiplets. Medusa Point will reportedly share 12-core chiplets with desktop, not hybrid aside from possible low power cores in the I/O die, intended to reduce idle power.
    Reply
  • Alpha_Lyrae
    usertests said:
    I don't think it's dropped. It's just the limit for the smaller socket (SP8). Zen 6 core count for the big socket wasn't specified.

    Videocardz lists 128-core Zen 6 for SP7. Where they are getting that info, or if it's a guess, I'm not sure. They also have 16-channel for SP7, 4/8-channel for SP8.
    If we assume 16 CCDs with 12-cores, SP7 should offer 192 Zen 6 "classic" cores along with 16-channel RAM via new IOD. Although, this will depend heavily on what type of packaging AMD ultimately move EPYC to, as costs are certainly going to increase. Strix Halo is InFO and Infinity Fabric IFOP is simply implemented as bunch-of-wires instead of SerDes. This is generally fine, as 2 CCDs will have trouble saturating a 2x32B (512-bit) link when RAM is 256-bit, though there are edge cases where you will hit a physical limit.

    In EPYC Venice, this will need to be matched to a 1024-bit memory subsystem, which could end up as 8x32B links in total or maybe something higher in bandwidth to support intermediate caches (bandwidth amplifiers) across the package to accelerate CCD-CCD teaming/workload sharing. Think of it like RDNA's workgroup processor where 2 CUs can act as 1 WGP, but instead 2 CCDs act as 1 WGD with an intermediate cache joining them underneath. I don't think we're quite there yet, but eventually, we'll see something like this.
    Reply
  • usertests
    Alpha_Lyrae said:
    If we assume 16 CCDs with 12-cores, SP7 should offer 192 Zen 6 "classic" cores along with 16-channel RAM via new IOD.
    Wow, I forgot that to check that 128 isn't divisible by 12, lol.
    Reply