What Open Sourcing Power’s ISA Means For IBM i Shops
August 26, 2019 Timothy Prickett Morgan
Due to a conflict with a prior engagement at the Hot Chips conference, where IBM unveiled some of the aspects of the impending Power9′ processor that will prototype some ideas about memory subsystems that will appear in the future Power10 chips, we were not able to attend the OpenPower Foundation’s developer summit in San Diego last week. But IBM kept us in the loop, and we were intrigued to learn that Big Blue was open sourcing the instruction set architecture, or ISA, of the Power processors.
This step is perhaps an inevitable one, given Big Blue’s desire to make the Power chip more attractive in the datacenter as both a processor and an accelerator and to make it better connect with all manner of memories, storage, and accelerators. IBM did not go all the way and open source the finished designs, coded up in RTL format, of the Power8 and Power9 families of processors, and as far as we know, it has no plans to do so with Power10 or Power11. Such a move would enable anyone to grab the designs – really a mapping out of transistors and connectors on their many layers of etching – and create a set of masks at a foundry and print out their own copies of IBM’s chips.
If IBM did that, customers could take the RTL code, suck it into chip design programs from Cadence Design Systems, Synopsis, or Mentor Graphics, make changes to the design, and then print them out at a foundry. In this case, that would be using the 22 nanometer or 14 nanometer processes at GlobalFoundries, where the Power8 and Power9 chips were manufactured, but Suzhou PowerCore, a PowerPC chip manufacturer in China, has licensed the full Power8 and Power9 chip designs from IBM and has been etching modified processors and we understand that the idea was to eventually make them at Semiconductor Manufacturing International Corp, the upstart foundry in China that wants to compete with the likes of GlobalFoundries, Intel, Samsung, and Taiwan Semiconductor Manufacturing Corp. The latter is arguably the most advanced foundry in the world, but Samsung, which is IBM’s fab partner for Power10 and beyond, is no slouch either. Intel has had woes with its 10 nanometer processes, which are akin to the 7 nanometer processes in which TSMC is leading today, and GlobalFoundries just backed off completely and stopped making massive investments in its move to 7 nanometers last year because it was not going to be able to keep pace.
What IBM has done is open sourced the instruction set architecture for the Power chips, which includes the PowerPC AS instructions that have the special memory tags in it that allows for IBM i and its single level storage to run on the architecture. The opening up of the Power ISA allows companies to create their own designs, but they have to figure out how to lay out the transistors to provide those functions, and they may implement them in a different way from how IBM created it. Much as AMD has a license to the X86 architecture created by Intel – one that came through black box, reverse engineering Intel’s processors for a few decades and hard-fought legal battles to be able to keep that intellectual property. Now, anyone who wants to create a Power chip can do so.
“We started OpenPower six years ago because the industry was seeing the decline of Moore’s Law, and we were seeing the need for more powerful systems to support HPC, artificial intelligence, and data analytics,” says Ken King, general manager of OpenPower at IBM. “We needed to find other ways to drive system performance, and with limitations on the processor, the ability to integrate and innovate up and down the stack was becoming more critical. This led to things like NVLink with Nvidia, a close relationship with Mellanox on interconnects, and OpenCAPI for other devices, and we have seen some progress here. But we are also seeing a shift in the industry, with companies moving to more open hardware. IBM opening up Power to the point where we would license the CPU RTL to others so they could design their own processors was limited in its effect because there were not that many people who wanted to spend many hundreds of millions of dollars – not for license fees, but for full development – to create their own high-end CPU.”
So, if you want to create your own processor to run IBM i and run it really well – and I shared some ideas about how I might do this a few weeks ago – now is your chance. My strong advice before you do that, though, is to figure out the legality of running IBM i licenses on non-IBM hardware, because that is a lawsuit just waiting to happen. But, the point is, innovation outside of IBM is possible and equally importantly, there is a chance to keep IBM i alive above and beyond IBM’s own commitment to the platform should it come to that.
There are some caveats to using that Power ISA, which is being opened up at the same time that the OpenPower Foundation is moving itself under the umbrella of the Linux Foundation, the place where Linux and a lot of cloud-native bits of software (including the Xen hypervisor) all live. It is also where Linus Torvalds gets his paycheck from, as do a bunch of other key open source software developers.
The Power ISA is being opened up in a way that allows for companies making chips can do so without having to pay the Linux Foundation, the OpenPower Foundation, or IBM any royalties. IBM, King explains, is also going to move the governance model of the Power ISA to a board at the OpenPower Foundation, and IBM will have exactly the same one vote that other members of that board has and no more. IBM reserves the right, of course, to make any changes it wants for its own platforms to the Power ISA, and by definition going forward those changes will be complaint with the Power ISA standard, but with this change in governance, and change that remains compliant with the standard can be done with a majority vote. If a non-complaint change is made – meaning something that would break current compatibility with the Power ISA, which IBM has maintained faithfully and extended since 1990 – it requires a unanimous vote of the OpenPower governance board members. That means IBM can cast the deciding vote if something needs to be done that breaks compatibility.
Compliance with the architecture is a big deal. AMD chips have to run software compiled to run on Intel chips, period. Software compiled for one Arm chip has to run on another Arm chip, period. So if IBM is a hardass about compliance, this is a very good thing.
The only thing we care about in the IBM i base is maintaining support those special memory tags that let single level storage work, but in theory if these are a pain to implement, the Power community could decide to rip that out. This seems extremely highly unlikely, and even if that did happen, a single Power8, Power9, or Power10 chip would be more than enough compute capacity for 95 percent of the IBM i base as it currently uses capacity.
If IBM dropped the price on IBM i software and on the capacity running it and also put an IBM i shell over Linux partitions, creating a kind of Linux PASE, we could see IBM i adoption broaden radically. But IBM might not bring in any more net revenue and the IBM i business might be less profitable, even if it was healthier. IBM has done this very thing with Linux on System z, so it is amazing that the same strategy has not been applied literally to the IBM i base. Go figure.
The other important thing is that companies that grab the Power ISA and innovate to create their own chips have patent rights to whatever they create and also they are backed, to the extent that a patent is applicable, by IBM’s own substantial patent portfolio. To help customers along, IBM has created softcore version of the Power ISA that is running atop Xilinx FPGAs, which will help to kickstart Power chip development for those who want to dive in. IBM is more keen on fostering development of accelerators and embedded devices based on the Power ISA than on data center-class processors, but I still think it would be interesting to create overclocked Power processors with very few cores and lots of cache and memory bandwidth that would be a real screamer on IBM i transaction processing workloads. If it didn’t cost $500 million to bring a server class chip into being, and if IBM would allow IBM i to run on machines equipped with such a chip, I would think about doing such a thing myself.
Maybe we can do it anyway? I know a bunch of venture capitalists. I know good people at IBM. We could give Big Blue a cut of the action, and passthrough the operating system and system software sales and get a cut of that action sent back to us to help fund the ongoing development cycle. Anyone want to work for the Advanced System Company? We could get an address at 400 Rochester Street, in Winston Salem, North Carolina, to have our headquarters.
Anyone want to play?