AMD Next-Gen EPYC Venice Zen 6 CPUs Reportedly Drop Into New SP7 Platform

News
By
published

AMD SP7 platform may have up to 16 memory channels.

AMD
(Image credit: AMD)

AMD's 6th Generation EPYC processors codenamed Venice and based on Zen 6 cores, will reportedly use an all-new SP7 socket, according to renowned leaker YuuKi_AnS, who seems to have seen a roadmap of a server maker. The new platform will enable support for up to 16 memory channels, he asserts, which will ensure enough memory bandwidth for CPUs with hundreds of cores.

AMD's 5th Generation EPYC processors codenamed Turin, will keep using the existing SP5 socket, which aligns with AMD's strategy to use one socket for two generations of server CPUs. But its successor — the 6th Generation EPYC processor codenamed Venice with Zen 6 cores — will use its all-new SP7 platform. The information is strictly unofficial, as AMD has not discussed its data center platform in several years.

The new CPUs will support 12 or 16 memory channels supporting DDR5 and innovative memory modules like MR-DIMM and MCR-DIMM to feed Zen 6 cores. While the number of cores supported by the processor is unknown, we can expect it to increase dramatically from 96 – 128 cores in the case of SP5 processors.

Given the core count increase, we expect the new SP7 socket to boost the maximum power supply for Venice processors. The Socket SP5 can provide a peak power of up to 700W, which is good enough for current CPUs, but for AMD's 6th Generation EPYC CPUs, that figure will likely increase significantly. Given that market players like Intel and TSMC are experimenting with various innovative cooling systems for upcoming multi-chip solutions that will consume and dissipate well over 1 kW (1,000W) of power, we would expect AMD's CPUs in 2025 – 2026 to be in the same ballpark.

What remains to be seen is how many pins SP7 processors will have. SP5 has 6,096 pins, so it is reasonable to assume that CPUs with up to 16 memory channels and higher power consumption will likely use considerably more. Meanwhile, a larger CPU package will allow AMD to place more chiplets on the processor package to increase the core count and functionality.

AMD has not publicly talked about its 6th Generation EPYC processors, so take the leaked information with a grain of salt, as even if it is accurate for now, AMD may change its plans.

Anton Shilov
Anton Shilov
Freelance News Writer

Anton Shilov is a Freelance News Writer at Tom’s Hardware US. Over the past couple of decades, he has covered everything from CPUs and GPUs to supercomputers and from modern process technologies and latest fab tools to high-tech industry trends.

See more CPUs News
11 Comments Comment from the forums
  • bit_user
    Given that market players like Intel and TSMC are experimenting with various innovative cooling systems for upcoming multi-chip solutions that will consume and dissipate well over 1 kW (1,000W) of power, we would expect AMD's CPUs in 2025 – 2026 to be in the same ballpark.
    Something tells me their power efficiency is going to start putting them at a serious disadvantage, relative to ARM, if they let power consumption keep running away to such an extent.

    Amazon's Graviton 3 uses just 100 W for 64 cores, for instance.
    Reply
  • edzieba
    bit_user said:
    Something tells me their power efficiency is going to start putting them at a serious disadvantage, relative to ARM, if they let power consumption keep running away to such an extent.

    Amazon's Graviton 3 uses just 100 W for 64 cores, for instance.
    Measuring relative performance merely by number of cores is about as useful as measuring GPU performance by number of fans.
    Reply
  • bit_user
    edzieba said:
    Measuring relative performance merely by number of cores is about as useful as measuring GPU performance by number of fans.
    I disagree in that it's a modern, 64-core server CPU with 8-way superscalar cores & 2x 256-bit vector FP + 8-channel DDR5 + PCIe 5.0. That should give you a ballpark idea of where performance is at. For a more detailed comparison, people can go and find Graviton 3 benchmarks - or, they can even run their own!

    The power data, on the other hand, is less readily available. Here's my source on it, BTW:
    https://www.semianalysis.com/p/amazon-graviton-3-uses-chiplets-and
    While I'm at it, here are some benchmarks. If you want better ones, find them (or run them), yourself:
    https://www.phoronix.com/review/graviton3-amd-intel
    There was a follow-on exploration of compiler tuning:
    https://www.phoronix.com/review/graviton3-sve-gcc12
    Reply
  • edzieba
    bit_user said:
    I disagree in that it's a modern, 64-core server CPU with 8-way superscalar cores & 2x 256-bit vector FP + 8-channel DDR5 + PCIe 5.0. That should give you a ballpark idea of where performance is at. For a more detailed comparison, people can go and find Graviton 3 benchmarks - or, they can even run their own!

    The power data, on the other hand, is less readily available. Here's my source on it, BTW:
    https://www.semianalysis.com/p/amazon-graviton-3-uses-chiplets-and
    While I'm at it, here are some benchmarks. If you want better ones, find them (or run them), yourself:
    https://www.phoronix.com/review/graviton3-amd-intel
    There was a follow-on exploration of compiler tuning:
    https://www.phoronix.com/review/graviton3-sve-gcc12
    I would recommend looking at the benchmarks you linked, as their comparison (16 core Graviton3 isntance v. 8 core Xeon instance) can vary wildly in comparative performance, let alone core-for-core performance, depending on the benchmark or workload.

    Comparing core counts is a useful metric solely when comparing within the same architecture. Comparing between architectures is it effectively worthless. "More cores is more better" is great for marketing copy, but not much else.
    Reply
  • bit_user
    edzieba said:
    I would recommend looking at the benchmarks you linked, as their comparison (16 core Graviton3 isntance v. 8 core Xeon instance) can vary wildly in comparative performance, let alone core-for-core performance, depending on the benchmark or workload.

    Comparing core counts is a useful metric solely when comparing within the same architecture. Comparing between architectures is it effectively worthless. "More cores is more better" is great for marketing copy, but not much else.
    You're missing the point. A bit more or less performance per vCPU here or there doesn't change the fact that Graviton 3 has way better perf/W, which is all I was really trying to say.

    This isn't directly relevant, because they quote whole-chip power but only rate-1 SPECint, but you can get some idea of IPC from it, if you normalize by clock speed:


    Source: https://www.tomshardware.com/news/jim-keller-shares-zen5-performance-projections-kind-of
    Note: you might have to view the article version. I tried embedding what it linked as the source, but it didn't render in my forum post.
    Reply
  • edzieba
    bit_user said:
    You're missing the point. A bit more or less performance per vCPU here or there doesn't change the fact that Graviton 3 has way better perf/W, which is all I was really trying to say.

    This isn't directly relevant, because they quote whole-chip power but only rate-1 SPECint, but you can get some idea of IPC from it, if you normalize by clock speed:

    Source: https://www.tomshardware.com/news/jim-keller-shares-zen5-performance-projections-kind-of
    Note: you might have to view the article version. I tried embedding what it linked as the source, but it didn't render in my forum post.
    No, you're just looking only for what you want to see.
    From your previous link, here's a workload where Gravtion3 pulls out well ahead (1.5x perf/core vs. the Xeon):
    And here's one where graviton3 is utterly crushed (9.9x perf/core for the Xeon vs. Graviton3):

    Workstation and server performance is HIGHLY workload dependant. The optimal architecture and CPU will vary wildly depending on what you are actually doing with the system. 'More cores is more better' is ignorant for consumer use, but outright incompetent for server and workstations.
    Reply
  • bit_user
    edzieba said:
    And here's one where graviton3 is utterly crushed (9.9x perf/core for the Xeon vs. Graviton3):

    Workstation and server performance is HIGHLY workload dependant. The optimal architecture and CPU will vary wildly depending on what you are actually doing with the system.
    That's not a simple matter of fundamental CPU ISA performance or micro-architecture. There's clearly something else going on there, which would need to be understood. Most likely, it some kind of x86-optimized code path that the test is stressing.

    Unless delving into the particulars of that test case, such outliers should be trimmed.

    edzieba said:
    'More cores is more better' is ignorant for consumer use, but outright incompetent for server and workstations.
    Again, you're not seeing the forest for the trees. It's not about which CPU has better performance, but rather about perf/W.

    Performance per vCPU is in the same ballpark, meaning Graviton 3 is a viable option for many or most workloads. It's not only cheaper because Amazon doens't have to pay the overhead of buying the CPUs from AMD or Intel, but also because they're cheaper to run. Better perf/W means lower cooling and direct power requirements.

    If AMD and Intel start selling mainstream CPUs that burn 1 kW, they're really going to feel a backlash from the industry.
    Reply
  • edzieba
    bit_user said:
    If AMD and Intel start selling mainstream CPUs that burn 1 kW, they're really going to feel a backlash from the industry.
    It's industry that are demanding ever-increasing performance density. It's no good having a 100W chip if you need 10 of them to do the work of a 1000w chip, whilst also taking up more rack space and paying 20x in per-core licensing fees.
    Plus, with no independent benchmarking of actual Graviton3 hardware, the perf/watt claims from Amazon are best taken with a large fistful of salt (even if you have low-level but not physical access, reported figures can be quite far off actual draw).
    Reply
  • bit_user
    edzieba said:
    It's industry that are demanding ever-increasing performance density. It's no good having a 100W chip if you need 10 of them to do the work of a 1000w chip,
    I think you missed the point about perf/W. Having better perf/W means like 2x or 3x 100 W CPUs doing the work of a 1 kW x86 one. That's a hypothetical example, but then so is yours.

    edzieba said:
    whilst also taking up more rack space
    The funny thing is that bigger power supplies and cooling solutions also take up space. So, a modestly higher number of significantly more efficient CPUs could actually come at comparable (or better) compute density. Amazon has 3x Graviton 3 CPUs per motherboard:
    https://www.servethehome.com/amazon-aws-graviton3-hits-ga-with-3-sockets-per-motherboard-designs-tri-socket-arm/
    edzieba said:
    and paying 20x in per-core licensing fees.
    I seriously doubt most of the software running in the cloud has per-core licensing fees.

    edzieba said:
    Plus, with no independent benchmarking of actual Graviton3 hardware, the perf/watt claims from Amazon are best taken with a large fistful of salt (even if you have low-level but not physical access, reported figures can be quite far off actual draw).
    You're confusing mass market consumer garbage with servers. Consumer hardware can basically do whatever it wants, as long as it doesn't fry itself. Servers strictly adhere to their power envelopes, period. If a server uses more power than it's supposed to, that's going to get escalated as a critical issue by the biggest of big customers, who will be hopping mad until it's fixed. That's because not only does Watts = $$$, for those folks, but if you exceed the specs to which their cooling systems were engineered, it can lead to a lot of badness.
    Reply
  • edzieba
    bit_user said:
    I seriously doubt most of the software running in the cloud has per-core licensing fees.
    Then you seriously doubt the majority of the HPC market.
    You're confusing mass market consumer garbage with servers. Consumer hardware can basically do whatever it wants, as long as it doesn't fry itself. Servers strictly adhere to their power envelopes, period. If a server uses more power than it's supposed to, that's going to get escalated as a critical issue by the biggest of big customers, who will be hopping mad until it's fixed. That's because not only does Watts = $$$, for those folks, but if you exceed the specs to which their cooling systems were engineered, it can lead to a lot of badness.
    No, you're confusing hardware sales with an Amazon-proprietary system. Since Graviton3 does not exist outside of Amazon-owned and Amazon-operated datacentres, Amazon can claim that a chip operates under whatever power budget they imagine, and nobody can independently verify that power consumption. Amazon can declare their '100W' figure is an absolute maximum power-virus package draw, or an 'average' draw, or a 'representative workload' draw, etc.
    As for other server chips that anyone can hook power meters up to, the story also isn't as simple as there being a singular power draw value. This has not been true for many years, as individual cores will increase and decrease their clock speeds depending on thermal headroom (and not just per-package, but per unit die area within a package), packages will increase and decrease power budget allocations depending on chassis budget, etc. Thermal management is an integrated system-of-systems, not a bunch of independent units with a '#never exceed' limit and a big specced-to-maximum-possible-load cooling system. Nobody builds out a system that dumb anymore because its a huge waste of money. We're over a decade from the Open Compute Project demonstrating that there are more knobs to turn than just 'slap on a power limit' when it comes to optimising DC architectures for efficiency (from physical component layout to power distribution to non-phase-change cooling architectures, etc).
    Reply