Google introduces POWER9-based Zaius P9 server design

Cloud infrastructures require hardware. The buck can only be passed so far before someone has to own and maintain a server. Google Cloud has announced its plans for the Zaius P9, an Open Compute Project server based on IBM's POWER9 CPU. The server design is chock-full of next-generation buzzwords like PCIe 4.0, OpenCAPI, 48V-POL, and 48v OpenRack 2.0.

The Zaius P9 server design sports a pair of IBM POWER9 CPUs and DDR4 memory. Google says the design draws on its twelve years of experience with the Open Server Specification, and improves upon it with updates for heterogeneous architectures and increased power efficiency.

PCIe 4.0 is easy enough for readers of The Tech Report to identify as the successor to the PCIe 3.0 technology present in modern desktop computers. OpenCAPI (Open Coherent Accelerator Processor Interface) is an extension of IBM's CAPI technology which allows for fast, coherent interconnect between CPUs and heterogenous compute resources such as ASICS, GPUs, and FPGAs. AMD, IBM, Google, Xilinx, and Micron are some of the members of the OpenCAPI Consortium. Notably absent is Nvidia, which is invested in its proprietary NVLink interconnect. 48V-POL and 48v OpenRack 2.0 refer to the replacement of the standard 12V power design in servers with a 48V replacement in order to reduce conversion losses and improve power efficiency.

Comments closed
    • Kaotik
    • 3 years ago

    Actually NVIDIA is part of OpenCAPI, but they’re at Contributor level with Xilinx, unlike IBM, AMD, Google, Mellanox and Micron who are at Founder/Board level

    • Anonymous Coward
    • 3 years ago

    Anyone want to join me in noticing the significance of Intel being beaten by IBM? First there was ARM at the bottom, now PPC at the top is actually taking over part of the market.

      • NTMBK
      • 3 years ago

      Or Google is using them as leverage to get better pricing from Intel.

        • Anonymous Coward
        • 3 years ago

        No, this isn’t the first time I’ve been reading this, there has been serious activity around Power8 as well. I am under the impression that Google is running a significant amount of Power8 at this moment.

        IBM has in fact made a product that is better suited to what Google wants to do, and Google can switch platforms at the drop of a hat, so they will.

      • the
      • 3 years ago

      IBM never left the top of the market. The change is that IBM is offering their high end chips to outside parties to produce servers optimized for their workloads. This also means that the 3rd party system designers don’t get AIX or IBM i (OS/400) which remains exclusive to IBM’s own systems.

      IBM’s efforts as a chip merchant has been rather fruitful considering that they’re one of the few companies that can challenge Intel on raw performance.

        • Anonymous Coward
        • 3 years ago

        Ah, but this is PPC beating x86 in the 2-socket realm, taking jobs once had by Xeon. This is not some flavor of big iron.

        Also I’m pretty disappointed by the downvotes. This is a sign of big changes in the processor market. IBM isn’t using better manufacturing tech than AMD can, they don’t have anywhere near the engineering staff that Intel has, yet they are beating Intel head to head. They are beating Intel on performance per watt, performance per dollar, [i<]beating Intel in a fair fight[/i<]. What sort of insecure person responds to this situation by downvoting?

          • the
          • 3 years ago

          POWER8 didn’t really have a dual socket version: they just had two different dies that was from single socket all the way up to 16 sockets. Another telling feature is that all POWER8 systems are using Centuar memory buffers, a distinctly highend feature. IBM has been over provisioning the lowend with the high end POWER8 chips because there isn’t a cheaper design targeted toward that segment. IBM selling POWER8 at the bottom is a long term plan to boost OpenPOWER market share. (And that’s not to be critical of IBM, I think their strategy had been a success.)

          POWER9 is the big news as IBM actually has a socket config that directly targets the dual socket market. This is truly targeted to compete against dual socket Xeon E5’s on performance and cost.

            • Anonymous Coward
            • 3 years ago

            The way IBM is gaining traction by, as far as I can tell, offering massive memory bandwidth makes me wonder if Intel will be forced to respond properly. I wonder if we will see a new server-to-desktop product for enthusiasts, with an insane L4 cache and memory config? 12 cores or whatever is a bit boring in comparison.

            I wonder how the extra SMT threads work out for efficiency on Power8 and Power9 compared to Intel.

            • the
            • 3 years ago

            AnandTech did some testing with the POWER8 regarding SMT. Their results indicated that POWER8 could be faster but it needed the higher SMT settings to pull away from the Intel chips in aggregate throughput. It was clear that Intel was winning in singlethreaded performance.

            I do think we’ll be seeing more and more chips with low core counts and high cache counts. Memory technologies like 3D Xpoint conceptually can replace DRAM with there per byte addressability but performance suffer. Large L3/L4 caches would help mitigate such a change. In fact with POWER9, only the scaleup designs will have L4 caches as they’ll continue to use Centuar chips. The dual socket scale out designs will be dependent upon L3 cache but have the benefit of directly connected DIMMs to the socket. I’m really curious how latency will play out with Centuar + L4 cache vs. directly connected DIMMs.

            On the Intel side of things, their next big server socket is designed to house big dies or numerous smaller dies on the package. The idea of L4 eDRAM cache has been floated around as speculation for that design choice early on. (Intel will be offering chips with FPGAs and Omnipath fabric so the speculation has settled on that as primary reasoning now.)

            • Anonymous Coward
            • 3 years ago

            I’m a bit skeptical of mixing small and large dies in a server environment, it seems to me that agressive SMT on uniform dies is superior because the processes on any given server are likely to be fairly uniform. (In my world view anyway.)

            Kinda don’t have time to write more about that now…

            Need to go read that Anandtech article too.

            • the
            • 3 years ago

            I think IBM is just going to have a common socket IO between the two dies types. I don’t think any system integrator will offer a mix of SMT4 and SMT8 dies on the same system. Conceptually they could as the two types have the exact same processor feature set and likely have the same coherency protocols for caching. I just don’t think anyone would want such a heterogenous config for the markets IBM is targeting.

            I do see some merit if offering the two core designs though. Some applications simply need higher core counts and others are burdened by single threaded performance. Businesses also have to pay for software so keeping raw core count low helps keep software licensing fees low too (thankfully few software companies charge extra for enabling SMT).

            • Anonymous Coward
            • 3 years ago

            I see on re-reading that you didn’t mean Intel wanted both types of core on the same die. Makes sense to me then. Both fat and thin cores have a place, I agree. I do my development on AWS and definitely choose the type of machine which offers the CPU, memory, storage and networking which is most appropriate for the task at hand. Quite a mix of machines at work in our account. Often there is too much CPU power bundled with the amount of IO we want.

            I don’t know what the ratio is, but I think AWS is renting more hardware to Linux users than Windows. Probably a significant part of their customers could switch processing architecture at any time, if there was economic advantage to be gained.

            Was a bit disappointed that AMD appears to have dropped the Bobcat lineage except what Zen might have borrowed.

    • cygnus1
    • 3 years ago

    [quote<]48V-POL and 48v OpenRack 2.0 refer to the replacement of the standard 12V power design in servers with a 48V replacement in order to reduce conversion losses and improve power efficiency. [/quote<] This stuff is pretty cool. It's taking more and more parts out of the traditional server and making them common infrastructure for an entire rack. They're basically taking the concept of a blade chassis and applying it to full size servers with the rack essentially being the chassis. In this case, it's a shared power supply. No need for every server to have a full PSU for AC to DC conversion if they can all just share one really big one that's more efficient. It also relocates and reduces the output of a heat source. And as cool as this kind of stuff is, you only get to work it with when you're dealing with thousands upon thousands of servers. edit: English blah

      • KeillRandor
      • 3 years ago

      So long as they’re reliable enough to balance out another single-point-of-failure for the entire rack, (rather than an individual server within it), then it’s cool, yes 😉

        • Pwnstar
        • 3 years ago

        Get a 2nd one for backup.

        • just brew it!
        • 3 years ago

        For high availability scenarios, I’m sure the rack would have dual-redundant PSUs with automatic failover (hot spare).

    • chuckula
    • 3 years ago

    I don’t know about Nvidia’s presence/absence from OpenCAPI but the pretty block diagram from the linked Google blog goes out of its way to show Nvlink as an alternative to OpenCAPI: [url<]https://4.bp.blogspot.com/-7tmbj422giE/WAEX3yE8VWI/AAAAAAAADL0/ENhrCuYFzXAilTv6n0DSUcm3X8iLDOA_wCLcB/s1600/zaius-1.png[/url<]

    • Captain Ned
    • 3 years ago

    So, will the successor to Zaius be Zira or Cornelius?

      • UberGerbil
      • 3 years ago

      [url<]https://youtu.be/M2E1m90YSpA[/url<]

      • davidbowser
      • 3 years ago

      That’s DOCTOR Zaius to you pal. He didn’t go to 6 years of Dirty Ape Medical School for nothing.

        • UberGerbil
        • 3 years ago

        It would’ve been only 5 years, but he kept failing his oral exams.

          • demani
          • 3 years ago

          That’s bananas.

    • MathMan
    • 3 years ago

    According to Venturebeat, Nvidia is member of OpenCAPI as well:
    [url<]http://venturebeat.com/2016/10/14/amd-dell-emc-google-hpe-ibm-mellanox-micron-nvidia-xilinx-launch-opencapi-consortium/amp/[/url<] Edit: and they are listed on the OpenCAPI website as well: [url<]http://opencapi.org/membership/current-members/[/url<]

      • WhatMeWorry
      • 3 years ago

      Are they serious or just keeping tabs on the competition?

        • backwoods357
        • 3 years ago

        Probably just hedging their bets in case people reject their proprietary crap.

          • MathMan
          • 3 years ago

          Is it really crap when there’s no alternative?

          Because, right now, that’s pretty much how things are. NVLink is available today, whereas OpenCAPI will be available a year from now at best.

            • Visigoth
            • 3 years ago

            Agreed 100%. NVLink is lightyears ahead of anything *right now* and will be in the near future as well.

            So unless you, backwoods357, can design a better, faster interconnect, big iron will stick with this NVLink “crap”.

      • the
      • 3 years ago

      I’d predict that nVidia will migrate over to OpenCAPI, at least on the physical and link layers. The higher level layers may remain proprietary but it would enable host chip designs (ie IBM) to make a single interface for both nvLink and OpenCAPI.

Pin It on Pinterest

Share This