back to article EPYC leak! No, it's better than celeb noodz: AMD's forthcoming server CPU

Media site VideoCardz has leaked two AMD EPYC 7000 server CPU slides revealing core, thread and clock details. EPYC, the renamed Naples, is AMD's forthcoming server CPU design, to be followed by updates codenamed Rome and Milan. It features: Up to 32 Zen cores 8 DDR4 Channels/CPU – up to 2,666 MHz Up to 2TB Memory per CPU …

  1. khjohansen

    Deep pockets? "Intel not Intel"?

    I'd imagine Intel would like to keep AMD alive - to avoid any monopoly issues.

    Didn't Microsoft invest in Apple back in the day, for similar reasons?

    1. Anonymous Coward
      Anonymous Coward

      Re: Deep pockets? "Intel not Intel"?

      If you mean the out of court settlement for the wholesale lifting of Quicktime code into Windows Media Player and the attached agreement to continue developing Office for MacOS, then ... no, they didn't.

  2. rav

    "...real competition to the Intel elephant? That seems unlikely,..."

    Hmmm...

    ...that seems unlikely?

    Seriously! AMD's Zeppelin "chiplet" which is the 8 core 16 thread foundation for AMD's upcoming MCM micro-architecture and retails PACKAGED for about $445 in the Ryzen 1800x version. ANd AMD is enjoying huge yield number as well.

    Intel has no answer that at all.

    Zeppelin combined on an interposer is the Treadripper 16 core 32 thread beast that DELL's Alienware is releasing as it's top gaming rig come July 27 2017. Threadripperis expected to cost about $850.

    Intel has no answer to that. The best it has is i7-6950x which is a paltry 10 core's and costs $1500.

    Intel however promises to release i9 18 core but again it will cost $2000. i9 is a massive slab of silicon and yields will be in the loo. Intel also does not enjoy the use of certain patents by Gabriel H Loh and AMD. Google that and see. AMD has beaucoup patents for memory control and cpu + gpu and stacked memory all on an interposer. Intel has done absolutely no work with this concept.

    EPYC is built from either Zeppelin or similar microarchitecture, an MCM that promises to be at the very least 60% of Intel cost while providing about 40%+- more bandwidth and ABOUT the same TDP.

    ...and YOU SAY AMD IS NOT REAL COMPETITION?????????????

    What do you want? The Spanish Inquisition?

    1. Anonymous Coward
      Anonymous Coward

      Re: "...real competition to the Intel elephant? That seems unlikely,..."

      It's good to see AMD performance-competitive with Intel again.

      It will be interesting to see how Intel responds - the margins on server CPUs are huge so the response may be price cuts.

      Thus I'm more interested in single-socket throughput at the top of the range; that will indicate whether AMD is really shaking things up.

    2. P0l0nium

      Re: "...real competition to the Intel elephant? That seems unlikely,..."

      An alternative reading of the same facts would say:

      1) Zen is a core that is 20% down on both IPC and frequency vs Skylake.

      2) GF's yields suck and they can't (cheaply) build anything bigger than 200sqm.

      3) In a contest between a large die and an MCM - the monolithic solution always wins.

      4) You know nothing about Intel's yields (and neither do I , these days:-)

      I'm sure that YOUR needs can be met with an AMD processor.

  3. Nate Amsden

    disappointed

    I was optimistic these chips would be good. I guess (WAY)too optimistic. My current servers run dual E5-2699v4 which is 2.2ghz 22 core and 145W TDP. Also 24x16GB sticks of ram each.

    I was expecting the new AMD chips to be 32 core and at most be say 130W. Their lowest TDP of any chip is 120W. Compared to the 25W TDP of the 2.1G Xeon E3 1240L v5 quad core(8T) that I run in my personal server.

    AMD's high end 2 socket comparison is a single 32 core chip being faster than dual 10 core chips.

    I was hoping for something better to replace my aging DL385G7s with opteron 6176 and 6276. But after seeing this likely will stick to intel. Another 2 or 4k per server cost doesn't really matter when the servers plus software are over 30k each already.

    Not likely AMD will win much virtualization customers with these numbers. I imagine these AMD chips run too hot for most/all blades too.

    So yes, pretty disappointing.

    1. Michael Sanders

      Re: disappointed

      Yeah. But isn't that the 1.x GHz Xeon?

      1. Nate Amsden

        Re: disappointed

        Not sure what you mean by 1.x Ghz Xeon, in my OP both Xeons that I referenced were 2Ghz+.

        If AMD could make a 32-core 2Ghz chip @ 130W, and charge 20% more than Intel's high end part(s) I would buy it without hesitation. Their wattage numbers are just crazy, even their 16 core parts are 155W (up to 170W ?? based on the article - which is higher than Intel's 22 core Xeons). This just doesn't seem like a good enough attempt to break back into the data center market.

        AMD seems aiming squarely at the 1 socket market. I can certainly see some use cases for such systems, am not sure how much of the market they make up. Last time around (Opteron 6000) AMD argued you don't need 8 socket systems, when their chips can do 4 socket real cheap with lots of cores. Now they say you don't need 4 socket systems(which are a small part of the market I'm sure), but having a pair of 150-180W CPUs in a two socket system just seems crazy.

        I want AMD to do well.. but they abandoned the data center market for the past 4+ years, it will take a lot to come back, and these wattage numbers make me think it is not enough.

        1. Anonymous Coward
          Anonymous Coward

          Re: disappointed

          I think you misunderstand AMD's plan here.

          They are claiming that Intel are screwing over the market by crippling performance of single socket servers, and basically forcing people to buy dual socket servers just to meet their I/O requirements.

          AMD are trying to show that the EPYC platform can replace most of the low-end dual Xeon platforms with a single socket set-up that outperforms them. AMD published a graphic which shows the current distribution of the server market and how they believe it will look with EPYC.

          Here is a PDF from AMD's own site on the disruption plan for EPYC which explains what I briefly mentioned in more detail.

          http://www.amd.com/system/files/2017-05/TIRIAS-AMD-Single-Socket-Server.pdf

          Yes I am sad. I still geek out on server hardware, particularly potential game-changing hardware, even after all these years....

          1. P. Lee

            Re: disappointed

            >forcing people to buy dual socket servers just to meet their I/O requirements.

            How much SSD storage and how many 10/40G ports do you get in a box before the CPU just can't pull data in and push it out fast enough? What happens when you use put FreeNAS on the new chip? Software defined networking? Media servers?

            In these scenarios (unlike gaming or transcoding), you don't need massive FLOPS, you need I/O throughput.

            I'm not sure it makes sense to compare AMD and Intel and say one is "better." We probably need to ask which is more appropriate for a given application. Intel generally wins in CPU-bound operations, it looks like AMD will win on I/O.

    2. Anonymous Coward
      Anonymous Coward

      Re: disappointed

      Considering the slides are incomplete, currently they show a single E5-2699V4 22C/44T being beaten by 47% by a EPYC 7601 32C/64T. I accept we need the full deck to judge as most tests are cherry picked.

      You also get 2TB per socket versus 1.54TB total RAM and 55 MB SmartCache versus 3MB L1/16MB L2/64 MB L3. We are routinely seeing 32GB and 64GB VM's rolling into production now, and the lower core count parts will interest our DBA's who are currently using E5v3's/E5v4's with the maximum memory installed.

      You've pointed out AMD loses in TDP, but I'm interested to see why some of the parts are bracketed as 155W/170W - there maybe some hope for blades.

  4. Martin Gregorie

    What is the security feature of which you speak?

    Is this by any chance a case of AMD emulating Intel and providing a trapdoor that can bypass OS security? I think we should be told.

  5. Marco van de Voort

    7281 faster but cheaper than 7351IP

    7281 lower clocked, multi socketed and cheaper but faster than 7351IP?

    Or does it simply have larger caches, and a SPECINT is a biassed benchmark?

  6. Christian Berger

    Well ARM would need a single hardware platform

    Essentially an "ARM-PC" where you can boot any install media without having to specifically port your operating system to your SoC.

    Essentially they'd need to have some basic peripherals (like serial ports or SATA/SAS) interfaces defined and always on the same ports. Then they'd need to ditch UEFI for something much simpler (i.e. Open Firmware), to satisfy the demand for boot systems that are sufficiently bug free to not cause major problems.

    The niche for ARM would be cheap "single purpose" servers. Kinda like the blade systems of the past, but _much_ less expensive. Perhaps systems that load an OS image into RAM and then run a single service on it. ARM is good at scaling down the power of a system that just sits around idly.

  7. John Savard

    Can AMD be bought?

    If a purchase of AMD by another organization would not void the agreement under which AMD is allowed to use Intel patents, I would be inclined to suggest IBM. In fact, if IBM, AMD, and Apple were combined under a single ownership, the resulting firm might have a chance of holding its own against the Wintel duopoly.

    AMD would make x86 server, laptop, and desktop chips. And offer the same technology in alternate SKUs with the PowerPC and z/Architecture instruction sets instead, as instruction decoding makes up a small part of a CPU. Presto, no longer the problem that forced Apple to move to the x86, and so the reborn Amigas could safely stay with the PowerPC.

    Apple would have management that made sense, instead of emphasizing the "insane" part of "insanely great". However, the iOS part, as opposed to the Macintosh, should be left in a successful separate company, as the fit with IBM's core business would not be a good one.

    1. P0l0nium

      Re: Can AMD be bought?

      AMD loses their right to use X86 if they experience a "Change of Control" ... And that doesn't apply to the other party's use of AM64 (Intel).

      Only the entity that has a "Change of control" loses its license.

      It was done to prevent any large entity (e.g, a government) strong-arming their way into PCs.

      (Which has effectively now happened anyway with all these ARM smartphone SOCs being built in government sponsored 14nm Fabs).

  8. This post has been deleted by its author

  9. Mikel

    AMD acquisition speculation

    AMD's cross-licenses of Intel patents are void on a "change of control". There will be no such acquisition.

    A single socket EPyC looks like just the thing for a workstation build. Getting pretty excited for year end.

  10. Joerg

    AMD lies as usual...

    This tells everything about AMD marketing claims and babbling about Ryzen and the the EPYC version "We note that Zen has 128 PCIe lanes, a lot compared to Intel, but when you make a 2-socket Zen machine, 64 lanes are used by each to talk to each other, as the architecture is built on PCIe. "

    AMD lies more than most manufacturers. They keep telling lies and doubling their actual specs numbers. They are still trying to sell the bad Polaris GPU as better and faster than the Nvidia Pascal ones.

    1. Mikel

      Re: AMD lies as usual...

      The EPYC platform has more PCIe lanes than Intel's. Period. That is what matters.

      AMD is totally up front that some of the lanes in each socket are used for their Infinity interconnect. But they still net more lanes than Intel.

    2. Anonymous Coward
      Anonymous Coward

      Re: AMD lies as usual...

      Not AMD Lies, Joerg understanding fail :)

      When you put two chips together, they use 64 lanes each to talk to each other, and still have 64 lanes each for I/O so 128 lanes. So the EPYC platform, as stated, always has 128 lanes..

  11. Tony Ferrelli

    EPYC... where have I heard that name before?

    EPYC.... oh, yes. It reminds me of EPIC, the infallible architecture behind the Itanium processors.

    Cue Celine Dion hit song...

    1. Anonymous Coward
      Anonymous Coward

      Re: EPYC... where have I heard that name before?

      Why do think AMD came up with the name. It is a dig at Intel. :)

POST COMMENT House rules

Not a member of The Register? Create a new account here.

  • Enter your comment

  • Add an icon

Anonymous cowards cannot choose their icon

Other stories you might like