In a video on YouTube, which has since been hastily removed, Hewlett Packard Enterprise (HPE) opened the can a little early on announcing a new dual socket AMD EPYC based system in a 2U form factor. As pointed out over at ServeTheHome, this is an important metric in the EPYC story: it is one of the first machines from a top 3 server equipment manufacturer.

Before the video was removed, it showed a dual-socket design with a full set of memory slots (that’s 32x, supporting a total of 4TB). Leveraging the 128 PCIe lanes that the configuration would bring, the DL385 Gen10 showed support for up to 24 NVMe drives (or 30 2.5” SATA SSDs) as well as either three double-wide GPUs or five single-slot GPUs. The aim for such a server seems to be a crossover between storage and compute, or the ability to maintain constant compute throughput with plenty of memory and fast local storage, such as large datasets for AI or deep learning network training. Redundant power supplies and HPE iLO management are also featured.

In AMD’s press release for this week’s Supercomputing 17 event, HPE was listed as one of the vendors now ready to start offering the availability of EPYC-based systems, now that the major cloud and hosting providers are getting to grips with the technology. HPE VP and GM, Justin Hotard, was quoted in AMD’s press release, saying that ‘AMD delivers the power … to help break barriers’, but no official or specific products were mentioned in the quote given to AMD. Along those lines, we would expect HPE to have its own announcement, probably this week as it is SuperComputing, although posting/deposting a video says a lot.

We suspect that when HPE pull the proper trigger on the launch, the video will be re-enabled and data sheets will start flowing. We will report when it happens, but here’s the video embed in case it comes online soon.

Update 11/20

The HPE ProLiant DL385 Gen10 website is now live, with the focus on lowering the TCO per Virtual Machine. Shipping starts in December.

The HPE ProLiant DL385 delivers unmatched security via the HPE Silicon Root of Trust, a unique link between the HPE Integrated Light Out (iLO) silicon and the iLO firmware to ensure servers do not execute compromised firmware code. The HPE Silicon Root of Trust is connected to the AMD Secure Processor in the AMD EPYC™ SoC so that the AMD Secure Processor can validate the HPE firmware before the server is allowed to boot.

Related Reading

Source: ServeTheHome

Comments Locked


View All Comments

  • peevee - Tuesday, November 14, 2017 - link

    64 cores/128 threads + 24 NVMe drives sounds like an awesome VM server (up to 63 VMs easily) or Hadoop node. Given that even one EPYC is essentially NUMA it probably not the best for single-node computing.
  • mdriftmeyer - Tuesday, November 14, 2017 - link

    More like the beginning of a wonderful Render farm.
  • peevee - Wednesday, November 15, 2017 - link

    Why render on CPUs, especially on AMD CPUs?
  • msroadkill612 - Wednesday, November 15, 2017 - link

    cos w/ 128 lanes, that's a lot of gpuS and other resources.
  • lefty2 - Tuesday, November 14, 2017 - link

    Hi Ian,
    Can you say if Anandtech are planning to do a Epyc / Skylake SP review in the future?
  • Ian Cutress - Tuesday, November 14, 2017 - link

    We have our launch day article:

    I have a few EPYC CPUs here for some workstation analysis, when I can get the motherboard to POST. Haven't spent much time debugging it yet. It's one of a few (dozen) projects on the go
  • lefty2 - Tuesday, November 14, 2017 - link

    Yeah, that was a good article, but it was only preliminary results on a early BIOS that didn't support full speed memory. No else has done any real world database benchmarks on Epyc, or Skylake SP. It'd be interesting to see how the infinity fabric holds up under those conditions.
  • rahvin - Tuesday, November 14, 2017 - link

    How about you send me on of those 1P processors along with a motherboard. You still can't buy the bloody things.

    Either AMD is screwing up big time (no serious wafer production) or they are still selling everything they make to the Cloud Companies/OEMs. I suspect the first at this point given how few OEM's have them for sale. There's no way Facebook/Google/Amazon/Microsoft are still getting the whole production run if they've been producing wafers a full steam this whole time. I'm willing to bet when we get the 4Q report we'll find out they have had production trouble and haven't sold hardly anything.
  • willis936 - Tuesday, November 14, 2017 - link

    The video is unavailable.
  • MrCommunistGen - Tuesday, November 14, 2017 - link

    As stated in the first line of the article... Then of course it begs the question, why not just remove it?

    Regardless, I've got a buddy who will be excited about this (I'm excited too but he'll be EXCITED).

Log in

Don't have an account? Sign up now