The Prospects For A Power9 Revolution
August 22, 2016 Timothy Prickett Morgan
As a relative underdog as a supplier of processors in the datacenter now that it has left the X86 server market and has gone from being a partner of Intel to a direct competitor that is fostering its own ecosystem in the richest and most competitive parts of the market, IBM wants for the Power processor to be able to take on the Xeon chip and win. The Power9 chip from Big Blue has about the best chance of any chip we have seen from IBM since the Power4 way back in 2001 to do this.
The top chip engineers at IBM will be on hand at the Hot Chips 28 conference next week in Silicon Valley, just down the road from Intel’s headquarters, and we will be on hand to get all the details that Big Blue will divulge about the Power9 chips that are expected to be rolled out in 2017 for scale-out machines and in 2018 for larger SMP machines. We talked a bit about the Power processor roadmap, which IBM revealed at the OpenPower Summit back in April. As we have said since IBM started up the OpenPower consortium to try to get other system makers as well as hyperscalers and service providers that build their own machines to adopt the Power chip for at least some of their workloads, IBM had to put out a complete roadmap that went out many years to demonstrate that it was committed over the long haul to the Power chip and was not going to sell off the business as part of its ongoing reorganizations and transformations. And to its credit, IBM did just that.
For now, it looks like IBM will still be leading the Power chip development effort, with other partners, such as China’s Suzhou PowerCore, taking IBM’s designs and tweaking them for use inside of systems that will be sold in China and perhaps in other Asian countries. We do not expect for Suzhou PowerCore to get much traction in the United States and Europe, but stranger things have happened so we are not ruling it out. The main thing is that IBM is not in this Power fight alone anymore, and it has enlisted the help of Google, Nvidia, Mellanox Technologies, and Tyan initially and now hundreds of others (including Rackspace Hosting and its own SoftLayer cloud) to try to push adoption of the Power architecture. These efforts are exclusively devoted to systems that run Linux, as you might expect, and we think that while many have wanted Microsoft to report Windows Server to the Power chip since it first did so two decades ago, this is not going to happen. But, let’s reconsider this, just for fun.
The Power9 chip represents the first time in a decade and a half where the architecture and the desire for a Xeon alternative has never been higher or more pressing. The clamor for Windows on ARM processors continues, and for big analytics and database jobs, one can make a credible case for porting SQL Server to Power running on Linux. Microsoft is, after all, porting SQL Server to Linux on Xeon processors, so hell has already frozen over. And Microsoft’s Azure cloud might be compelled to use Power-based servers internally for services even if Microsoft never does productize it and set it loose on its vast Windows Server installed base directly. Microsoft could use it on Azure first–and do so for years–before productizing it in a distribution for private datacenter use. That is what we would do, just like Google and Rackspace Hosting are planning to do with their “Zaius” Power9 server.
As we have previously reported, IBM is creating two different flavors of Power9 chips. One is aimed at machines with one or two sockets, which it calls “scale out” machines, and another is designed for larger NUMA systems with four or more processor sockets sharing memory, which IBM calls “scale up” machines. The future chips are called Power9 SO, due in the second half of next year, and Power9 SU in the roadmap above. Interestingly, the Power9 SO machines will allow for normal DDR4 memory sticks to be connected to the processor complex, as is the case with Intel’s Xeon E5 processors, while the Power9 SO chips will have buffered memory circuits between the processor and the main memory to allow for greater memory bandwidth and capacity to be attached to the chips, much like Intel’s Xeon E7 chips.
An approach called co-design, which means tweaking hardware to better meet the needs of software, is becoming more and more popular among cloud builders, service providers, hyperscalers, HPC centers, and system builders, and we think that given the diverse customers that Big Blue is pursuing, it is very likely that the Power9 chip lineup will be considerably more diverse than the Power8 families before it. IBM will have variations of the Power9 SO and SU chips for its own Power Systems machines and tuned specifically for Linux, AIX, and IBM i workloads, and then there will be other versions aimed at so-called “merchant silicon” distribution, meaning chips that go into other machines that are either built by OEMs and ODMs, sometimes for their own business of pushing machines directly to customers and sometimes at the behest of customers that come to them and pull a machine through their factories. There is, of course, a limit on the customization that Big Blue and its foundry partner, Globalfoundries, can support and not lose money, but by the same token, there is a certain amount of variation and customization that must be made available to be in the game at all given this co-design approach.
Interestingly, such variation will give IBM i shops more options than they have had in the past–if Big Blue is encouraged to keep its IBM i options open. That means we need to do the encouraging, or else IBM i will be at the back of the line. While we understand that IBM needs to cultivate its Linux on Power business and that the vitality of Linux on Power directly correlates to the health and investment in the Power platform overall, we also think that any benefit that is given to Linux must be given to AIX and IBM i, too. Now that we are about a year out from the initial Power9 SO system launches, this is the time to make your voices heard to IBM, which is formulating its system plans now for next year. I have said it before and I will say it again. Any machine made by any ODM or OEM or even IBM itself that can run Linux, either on bare metal or on the OpenKVM hypervisor, should be enabled to support PowerVM and IBM i and AIX. I get plenty of criticism for thinking too much about hardware, but here is the deal: Without the hardware foundation, there is no IBM i platform. And every economic and technical benefit that comes to the Power architecture has to come to IBM i.
In a sense, IBM i is a special case of co-design, and we would argue one of the first instances of co-design starting back with the System/38 in 1980 and continuing with the AS/400 in 1988. For certain use cases, particularly with service providers building clouds, a machine designed to run Linux for hyperscalers would make a fine machine for MSPs building IBM i clouds where customers will likely have modest workloads and budgets. IBM didn’t seem to understand this two years ago, and maybe it will in 2017 when the Power9 SO machines come out.
No matter what, the new iron and its capabilities will be exciting, and some real competition is coming to the datacenter processor market, from IBM to AMD with its “Zen” Opterons to various ARM chip suppliers. So next year is going to be a lot of fun for hardware, and we will get our first glimpse next week at some of the innovation coming to bear in 2017 at Hot Chips. Stay tuned. And send us any comments on what you want to see from Power9 systems. Let’s start a feedback loop.