Powers Of Ten
May 11, 2020 Timothy Prickett Morgan
The one thing that is easy to predict about designing and manufacturing a server processor is that no matter who is doing it, no matter what process in what decade, it is always hard and things always run late. Whether or not we on the outside world can see this, it remains true just the same. And that is because in any given era, with any given chip etching technique or chip architecture, server processors are always pushing up to the very limits of what is possible. And things go wrong.
This is why it is a good idea to check in with server CPU makers from time to time to see how things are going, because unless something is going to cause a material change in the financials of a public company, they are not going to volunteer information about how things are going. They build in their roadmap buffers, fuzz up the lines a little bit, try to say as little as possible. If you are Intel or AMD, where server processing is in fact very material, then if something is going wrong you have to say something about it. Intel has been cagey about what has delayed its fourth generation “Ice Lake” Xeon SP processors, which were expected a few years ago and which are not coming out until later this year, but it has been explicit about the delays. It was similarly explicit, but subdued, when it quietly killed off the third generation “Cooper Lake” Xeon SP processors, which were due sometime this year but got bunched up against the Ice Lake Xeon SPs and therefore could be sacrificed. So now Ice Lake Xeon SPs will be the third generation, we suppose.
For the past six and a half years since Big Blue established the OpenPower Consortium with Google, Nvidia, Mellanox Technologies (part of Nvidia for a little more than a week now after a $6.9 billion acquisition closed), and Tyan, IBM has been remarkably explicit and consistent with the Power processor roadmap. Here is the one from a few years ago:
And here is the one from last year:
As you can see, the memory-enhanced Power9′ processor, which is aimed only at HPC and AI workloads, was pushed from 2019 into 2020, and in fact, we had expected it to come out already but maybe the coronavirus pandemic gave it a little more push out into the future. Or maybe IBM’s not getting the two big exascale system contracts at Oak Ridge National Laboratory and Lawrence Livermore National Laboratory – the future “Frontier” and “El Capitan” systems, weighing in at 1.5 exaflops and more than 2 exaflops, respectively – meant that IBM didn’t have to build prototype machines to give developers a taste of the high memory and I/O bandwidth architecture that was coming with Power10. We don’t know.
What we do know is that IBM also shifted the Power10 from being a “2020+” timeframe as in the first roadmap to “2021” in the second roadmap; we also note that IBM has also radically jacked up the memory bandwidth per socket with the Power9′ and Power10 chips between these two roadmaps, which is great news and which we discussed in detail in Power9 Prime Previews Future Power10 Memory Boost, published last September.
We think 2021 as a date is a little vague, and with Power9 entering the long tail of its generational cycle, with the pandemic putting a damper on Power9 system sales, and with IBM shifting fab partners from Globalfoundries with Power9 and Power9′ to Samsung for Power10, we wanted to check in with Big Blue to see what the current, most up-to-date plan was for delivering the Power10 chip. We were expecting Power10 in early 2021, which would have been approximately 40 months after the first Power9 chips rolled out in December 2017 aimed initially at HPC workloads and eventually coming to the commercial Power Systems platforms in early and middle 2018.
“Power10 has always been targeted at the middle of 2021 to later in 2021,” explains Steve Sibley, vice president and offering manager for the Cognitive Systems division. “There was certainly some speculation out there that it might come out earlier, and I don’t know if we fueled some of that with the earlier roadmaps. We said 2020+ in the earlier roadmap because at the time we didn’t want to commit to any year. From out schedule, Power10 has always been in the second half of 2021, and if you look at Power9 to Power10, that is actually what I will call a pretty normal schedule. Samsung is going reasonably well, and certainly we are seeing challenges right now because of COVID-19 because we can’t send all of the developers into the lab to debug the latest instance. We are seeing things take a bit longer to debug and get executed. So we may see a month or two slip from that and we might have to adjust our rollout plan slightly. But overall, the Power10 project is still going pretty well. So it is probably now fourth quarter instead of late third quarter. But it is not being pushed out because of demand, but rather because things are a bit more challenging to get done.”
In the scheme of things, this is a relatively minor delay in delivery of the Power10 chip, and we also think that if IBM had won those exascale deals, it may not have changed the timing one bit because those were also pushed out to a later date because of the difficulty of building these machines.
The real issue that we see is that it is a long time between Power9 and Power10, and there are a lot of Intel, AMD, Ampere Computing, and Marvell processors aimed at servers coming down to pike in the next two years, not to mention a few upstarts such as Tachyum and Nuvio. IBM is going to be delivering what looks like a killer advantage in memory and I/O capacity and bandwidth and acceptable compute with Power10, which will most likely have 48 cores, double that of the “Nimbus” scale-out version of Power9, or 24 cores, double that of the “Cumulus” scale-up version of Power9. We expect the threading model to stay the same: four threads per core on the scale-out version and eight threads per core on the scale-up version.
The question we have is what is Big Blue going to do to encourage customers to buy Power9 now and up until the middle of next year rather than just wait for Power10 entry machines in late 2021 or early 2022? This business has to generate revenues to pay for itself. So, hopefully, IBM has something up its sleeve other than just biding its time.