IBM has released the Power Architecture instruction set to the Linux Foundation
eejournal.com
eejournal.com
Wow. Why is it that when someone is confident in something and chooses to donate it to the people in the face of a changing market people treat them as if they've failed? This is a massive success and a new beginning - they've taken a closed system they've been working on for decades and opened it up - we should be applauding this action, not denigrating the effort.
https://www.bizjournals.com/albany/morning_call/2015/01/anal...
https://en.wikipedia.org/wiki/IBM_z14_(microprocessor)
Edit, Adding some microarch descriptions links below:
https://www.anandtech.com/show/11750/hot-chips-ibms-next-gen...
But thats like calling an x86 server an ARM because there are ARM's in the disk controllers, and on the motherboard acting as BMC's/etc.
(edit: misquote)
something that (x86) juggernaut can never do, going open
A trivial review of IBM's systems revenue versus the datacenter breakout of any other CPU maker earnings reports will net you: yes, it is indeed quite successful.
Revenue is a different story that doesn't necessarily reflect technical merit.
so, how are we judging cycles here? I seriously doubt Intel and AMD use the same definition that IBM do.
Let's just say between IBM's cycles:
- new PCI Express subsystem revision (first to market with PCI Express 4.0 in P9... who knows with P10)
- new memory interface architecture (Centaur and DDR4 direct attach w/ P9 and now OMI with P10)
- new coherent processor external accelerator bus (OpenCAPI and CAPI...)
- new internal interrupt accelerator function architecutre (XIVE w/ P9)
- new privilege level for hypervisor/guest separation (Ultravisor/PEF in P9+/P10)
... and so on.
Intel and AMD do NOT do that. Let's compare apples and apples here instead of using ill-defined words like 'cycles'.
As for the number of cores, memory bandwidth needs to keep those cores fed and now we're just seeing real memory bandwidth improvements - the number of cores and threads is expected to increase commensurately with that in the next revisions of the micro-architecture.
Shipping excellent, up to date, and competitive processors is more than just releasing a moderately tweaked, more tightly binned SKU every 6 months.
Neither AMD nor Intel shine though, if they were the target of the comparison. For the past 5+ years Intel more or less released slightly tweaked versions of the same CPUs and all those tiny bumps were just wiped out anyway so we can basically just count on marginal power improvements. And up until Zen in 2017 AMD had practically been a no-show for 6 years.
IBM was brilliant in how it rolled up the mainframe/old school Unix/AS400 business on the hardware side. But that playbook was based on big traditional companies and leveraging the base and relationships... there’s a reason that AIX essentially is the last proprietary Unix. But every EA renewal cycle Microsoft is shipping some Azure, some department is buying a SaaS platform, old school IT directors are retiring, etc. It’s a slowly dying business and has been for a long time.
Intel has its own problems, but they are a different category of problem. Even then, we have multiple suppliers, and Intels struggles ultimately help the x64 platform.
Performance and price never had a linear relationship. Past a certain point each "unit" of performance costs substantially more than the previous. Think 80-20 rule.
If you're running a big implementation of software that scales up or has licensing rules that make scale-out expensive (ie. Oracle), Power is a no-brainer. Problem is that generally speaking, the industry had gone a different path.
And probably it'll be RISC-V at the end of the day anyway, and it'll sell Redhat just fine. ;- )
I vaguely recall that it was a power hog.
I don't know where this idea comes from that open-sourcing a platform is a step down. Had Linux started out closed-source and then gone open-source, I think it would be clear to everyone (knowing what we know today) that this would be a huge benefit to all. Same with Javascript, or the JVM, or CLR, or any other platform. Open platforms at any level have more value, full stop.
Is keeping an ISA proprietary of real value to anyone these days? The cost of entry is still high enough that you're not going to be threatened by J Random Hacker. We can already see this. ARM is available for people to license, and TSMC will fab chips for you, yet Apple's CPU division doesn't seem to be seriously threatened by anyone.
IBM has poured a ton of money into having roughly the same ecosystem for POWER as x86: nearly every major Linux distribution ships ppc64le ports [0] (most at full parity to their x86 server release--including testing and building on the platform itself), there's several places to get access to it (even for free as I mentioned on a previous thread [1,2]), and several FPGAs, GPUs, and other accelerator cards work with it [3,4].
Going past x86 (and competing with RISC-V and others), there's a freely available softcore (as I mentioned elsewhere) [5]. And since IBM open sourced the latest ISA with a full, modern instruction set (AES, SHA, Vectors...) -- the community around a softore could eventually go and implement it [6]. And as Talos has shown, even most of the microcode is open source or at least publicly verifiable.
Once the community forms around it, what do you have? [7]
A starting point for all sorts of new projects. Someone else has done nearly all of the work to build something amazingly complex (from hardware gates to a working NodeJS/Go/$LANG stack): all you need to do is tailor this to what you need.
Want to optimize to your specific workload? Grab the softcore, pull out the portions you don't need, deploy/test on a FPGA, and run Linux over the top with a full suite of debugging tools already available.
Want to experiment with new instructions not currently available? Grab a softcore, implement your new instructions, add compiler support or assembly support, and iterate.
Want a low-power core for an IOT device? You could license ARM or you could build on a free POWER softcore.
Want tons of scalar cores for a massively parallel problem? Write it in C, compile to ppc64le, and deploy on many instances of the same softcore running on a FPGA.
None of this is possible with x86 or RISC-V today. If it is, the overhead, time, expertise, and cost makes it exclusive to only large companies and research institutions. IBM has done most of the work to commoditize this. Costs to fab will remain high but deploying on a FPGA would at least tell you if your approach is viable and something you could consider.
Yeah, Intel or AMD could eventually open source this much too. But they'd have to start having conversations with their lawyers... And then with all their partners... And then eventually they'd get to where POWER is today... :)
[0]: https://news.ycombinator.com/item?id=20752226
[1]: https://news.ycombinator.com/item?id=20749369
[2]: https://news.ycombinator.com/item?id=20751639
[3]: https://www.nextplatform.com/2017/12/15/nvlink-shines-power9...
[4]: https://www.servethehome.com/supermicro-p9dsu-c-based-ibm-po...
[5]: https://news.ycombinator.com/item?id=20826081
[6]: Yeah, there's admittedly a lot of work to go from "scalar softcore" to "100% ISA implementation". But if IBM does things correctly and tries to build a community, especially of academic researchers, around the softcore... It could go somewhere.
[7]: Ok, I guess this is an if... :)
All of the microcode and firmware is fully verifiable and owner buildable/modifiable. Even the secure boot root trust keys are publicly known and can be personalised with your own key with, of course, the well-known key distrusted. AFAIK, the only thing that isn't owner modifiable (or modifiable by anyone else really), besides the silicon, is the one-time programmable ROM which the processor bootstraps from on power-on is publicly verifiable w/ assembler source code. Datasheets and programming information is available for almost every single chip on the board - as are schematics (woohoo! I've already used them to make repairs a few times). As of right now, the only blob required, on the stock system, is related to the broadcom network chip firmware - and that's been fully reverse engineered and a beta version of the firmware which is clean room open source is now available (project ortega).
It's the most open and hacker-friendly system I've ever owned and I really wouldn't want to have to go back to something less.
To be fair, there's a bunch of small, decent quality, open source RISC-V cores that have gone to tape-out (RI5CY, Ibex, PicoRV32, SweRV)
I'm personally most interesting in profiling and tuning to my workload to run on FPGAs. But that's many years out. :)
Can a solo developer rent a cloud VM running POWER at a reasonable price yet?
If you're working on something commercial, you might check out http://integricloud.com/ -- looks like they start at start at $250/mo for a dedicated system (with 16-cores). If you check out their billing rates, it sounds like you can get a single core/thread though for about ~10-20 a month: http://integricloud.com/content/base/managevms.html?viewrate...
(I've never used integricloud but someone commented on my comment mentioning OSUOSL from a week ago)
It did. The original license didn't allow commercial use.
> “The future has never looked brighter for the Power architecture,” says Hugh Blemings, Executive Director of the OpenPower Foundation, apparently with a straight face. You don’t drop your price to zero when business is good. Still, the move probably is a good one for Power as an architecture. And a good one for designers the world over, now that they have a major new choice to consider.
> Your very own IBM computer. Who would’ve thought it?
I'm stunned. Who would write something as smugly as this is written... and why?
You don't reduce the price to zero when business is good, no. IBM isn't. They're reducing the price to zero for the parts of their product to which business is now irrelevant and for which free and open collaboration is now more valuable than secrecy and proprietary ownership - and good on them for it. They still charge lots of dollars for their support, their processors and their implementations based on them - and stand to benefit massively from opening the ISA and surrounding infrastructure and patents.
Do yourself a favour and don't write like the article's author does - if you do, it'll just make you sound like an ass.
Even the desktop "self-supported" version was something like $90 last time I checked.
CentOS/Scientific Linux is not RHEL.
RHEL 8 was released almost 4 months ago and there's no CentOS 8 yet. Mainly because taking the "free" RHEL sources and rebuild them isn't a simple case of removing logos and other "non-free" material. The build process itself (the sequence in which packages must be built so that all of them build successfully and with equivalent functionality to RHEL) is a major piece of what makes RHEL, RHEL, and that isn't free (or even publicly available).
And possibly bad for them. See Laws of Tech: Commoditize Your Complement [1], by Gwern.
> Joel Spolsky in 2002 identified ... the pattern of “"commoditizing your complement"”, an alternative to vertical integration, where companies seek to secure a chokepoint or quasi-monopoly in products composed of many necessary & sufficient layers by dominating one layer while fostering so much competition in another layer above or below its layer that no competing monopolist can emerge, prices are driven down to marginal costs elsewhere in the stack, total price drops & increases demand, and the majority of the consumer surplus of the final product can be diverted to the quasi-monopolist.
This move is good for IBM. Not clear what the repercussions will be for other players - even open source developers.
A good thing to point out is that while I have a vague notion that Rotschilds are bankers, I have strictly no idea what business the Rockfellers were in. I only know them because of their charity efforts.
Giving back is not a defeat, it is the culmination of success.
That you don't know what business Rockefeller was in would please him greatly beyond the grave. The man went to great lengths to obscure his business activities, and there was a concerted effort late in his life and by his descendants to ensure that people remembered the family for their charitable activities, and not the business activities that precipitated the largest anti-trust action in history.
Standard Oil - ExxonMobil, Chevron Corp., and BP own big parts of that business now.
And lots of other ventures, mines, real estate (remember this Rockefeller Center thing, plus half of NY), etc.
Enterprises are rarely giving something away, they should earn money, so if they do this, it rises suspicion that there is a product that has no value and a company gives away something worthless only for the sake of PR.
What is concerning here is that opening SPARC did not create a vibrant ecosystem of SPARC parts and computers, which is a shame.
As I see, IBM has the high end well covered with POWER9 and 10 and I'm not sure the embedded PowerPC crowd is that interested in big brother's ISA.
The key difference can be that while SPARC was stagnant when it was opened, POWER is extremely successful (and took over much of what was previously owned by SPARC and HP)
Compared to what? To the meteoric rise of x86 in the data center, and ARM in embedded, not enough it seems?
And yes, currently the two largest supercomputers (Sierra & Summit) are POWER machines, but 1) Most of the oomph in those come from the Nvidia GPU's not the POWER cpu's 2) IBM got none of the CORAL-2 systems, which makes people wonder whether they have given up on the HPC market.
If anything, it seems POWER is on a slow descent towards irrelevance. Considering the nice moves they have made wrt. openness (e.g. the Raptor machines with open firmware all the way down to the metal etc.), I hope I'm wrong and this move helps turn the ship around. We'll see I guess.
IIRC, the time Sierra and Summit got designed, POWER offered interconnects faster than anything available on x86, which was important to keep the GPUs in those systems busy. Different HPC systems have different needs and with the latest Xeon Scalable and AMD Rome that advantage may have evaporated a bit - and I'm not surprised IBM didn't win or even propose. Let's see what OpenCAPI and OMI bring to the table and if, with the good experiences with the Talos desktops as proof that the software ecosystem is there some foundry gets interested in something that could compete with lower end Xeon in the embedded space.
There are exceptions, though. Blender failed to make money for the studio that made it, resulting in them selling it to the community for a one-time fee, and it is now one of the great jewels of the open source movement: https://www.blender.org/foundation/history/
It's important to note, though, that in Blender's case the founder remained involved the whole way through, and I think still is. I suspect his dedication and long-term attention is a lot of why it worked out.
It's because of past experience. FOSS Foundations are where code/specs are donated to die.
Not because FOSS can't create success stories, but because when a for profit company gives away proprietary specs/code as open source, it's because they see no further benefit from keeping it, and its already dying. And of course, it's precisely the company that employed people working on the thing, and who know it well - a community wont emerge from nothing.
There are counter-examples but few and far between.
Less check back on this "great future" in 5 years...
Examine it first then make judgements on the product they are releasing not why they are doing it!
Did they release anything which wasn't publicly known before [1]? IBM "open sourced" the POWER ISA and gave patent cover, meaning they won't sue you if your create a POWER ISA compatible chip without paying them. But the ISA documentation was no secret before either; a secret ISA wouldn't be particularly useful after all.
[1] Yes, they released a microcontroller class softcore. Fine, but very far from POWER9.
It's a good move on behalf of IBM. But it might be a bit too late. If you were a company with a demand for custom silicon, you'd probably choose RISC-V as they already have experienced companies willing to provide support and services to meet your needs. With POWER your choices will be limited.
On it's face, yes - absolutely. IBM had to do it now if it didn't want its lunch entirely eaten out from under it. That said, that's a good thing. I think we'll see both architectures working from opposite sides and moving into each other's markets. It's still very early days for RISC-V and OpenPOWER, so hopefully they'll be able to benefit each other. We've got excellent open friendly hardware for OpenPOWER (Raptor Computing System's Talos/Blackbird and soon, Condor lines) and we're getting there with RISC-V embedded systems and mobiles.
At least now on both the high and low end of the performance spectrum we have open alternatives for all needs. Here's to a more open, more diverse future where none of us has to scramble to be everywhere, all the time, all at once.
If I cared about having an open AND mature ISA, I'd pick POWER now.
Since both RISC-V and POWER are now free, I'm expecting to see a mix of the best of both.
The PowerPC is perhaps better placed, may not get much farther.
But that's true of any project: throw it over the wall and abandon it, it will die.
As a bonus most of the firmware of the components on the board are open source as well, and just a git clone away.
(Not sponsored by them, but I've considered such a system for myself)
* https://www.phoronix.com/scan.php?page=article&item=rome-pow...
Not exactly int the “not at all” range imho.
1) SPARC was _never_ commercially competitive versus contemporaries. Not an ISA problem but they never had the design acumen other CPU makers did/do and the manufacturing partners always treated it second rate. I could stop here, there is simply no comparison or lesson in SPARC other than how not to run a business. It was successful in spite of this due to many other reasons (primarily ecosystem).
2) POWER9 is still competitive against chips just now launching 3 years later on a full step of manufacturing processes.
3) Other CPU vendors are about to hit an enormous memory wall. IBM is 4 years ahead of the industry here with OMI. Can still use cheap standard DDR4 by putting the buffers on board or on chiplets in P10. OMI submitted to JEDEC for standardization so it may be a total non-issue.
4) IBM's AXON PHY is 4 years ahead of everyone else in design and actually deployed. Glueless AI/ML/DSP nodes with GPUs or FPGAs versus i.e. nvidia DGX. Much better interconnect than infinityfabric.
5) P10, P11, and P12 are already planned. Expect future supercomputing wins especially with P11 (US gov likes to spread contracts among vendors every other generation). The enterprise POWER business is big enough to subsidize this game for a while.
If there is an error or impasse, it is that most developers now consume VMs on cloud services and for cloud operators the decision to move off x86 is a bureaucratic nightmare in the few companies that are technical enough to do it. POWER will live for a long time under IBM. The most likely outcome of this announcement is PowerISA MCUs and SoCs may see a resurgence. I can't see any trends clearly enough to predict POWER uptake outside that.
Also, take a look at the Protected Execution Facility (aka Ultravisor) ... it permits systems owners to separate hypervisor administration and code from guest machines and shrinks the trusted code base to almost nothing. Also, fully open source. Intel and AMD have something kind of similar but it requires proprietary firmware and trusted key signing by the vendor while IBM's design doesn't require any proprietary firmware or trusted key signing by the vendor at all. Very very cool.
They were far and away the market share leader for Unix server shipments in the late 90's and early 2000's.
You say they were the market share leader through the "early 2000's". In 2005, their OS was open-sourced. Sure enough, everyone I knew using Solaris since then has run it on x86. People like the OS but only used SPARC when they had no choice.
Sun was doing quite well with Sparc prior.
Right?
The ISA and implementation of POWER has been license-able from the beginning. You can still license implementations, depending on what you are doing that may be pretty cost effective.
You can't fab the same POWER9 that ships from IBM, but like you implied, someone could (eventually) fab a POWER9 CPU built on an open-source core.
The microwatt project [0] was demoed at the OpenPOWER Summit and is an open-source softcore. I'm very interested in watching to see what happens with this project and how much of the ISA they end up implementing. They've run a subset of micropython [1] on it and apparently gotten a few FPGAs working. The Summit talk [2] mentioned DRAM support and Linux support potentially being on the roadmap.
[0]: https://github.com/antonblanchard/microwatt
The only obstacle to fabbing it yourself would be that it's designed to be fabbed on GF's 14HP process, which incorporates some IBM-proprietary fab technology which was transferred from IBM when they spun off their own fab to GF; thus AFAIK the GF 14HP process is only licenced for IBM's use. Since IBM is willing to licence POWER9, perhaps they'd also be willing to negotiate access to the process.
(GF 14HP is different to GF 14LPP, which is what other GF 14nm customers (e.g. AMD Zen) use.)
But without actually opening any source...
Written initially by a Distinguished Engineer at IBM and sponsored by IBM, this was released when the announcement was made at the OpenPOWER Summit.
If IBM donating the Power instruction set and patents means that just one CPU will come out of it, although 50% more slow and costly than the competition but free from shady subsystems, ie a trustworthy CPU, I would call that a huge victory.
Closed chipsets device drivers are another similar problem, still that would be a good start nonetheless.
As an aside, it was long before Trumps trade war with China that USA refused to sell chips for Chinese supercomputers, prompting them to seriously start developing their own.
Will be interesting whether we'll see Chinese POWER designs, or whether they continue with MIPS and RISC-V.
Nothing works, until something does. And the something has little to do with the actual merits of the thing, but rather, the timing and confluence of events sureounding it.
As an example, the fact that a major Chinese company was fairly capriciously prevented from working may be a triggering event that would concentrate all Chinese companies’ minds towards adopting a new ISA and chip technology that’s open.
But to be competitive with Epyc, Xeon, and the Power 9 you'd have to spend a pretty huge amount of R&D. It does sound like there's some IP related to the memory controller which is available, but there's much required for a high end CPU that is not available.
What I do wonder is - what will they do with AIX? As I do foresee it going the way of Solaris, and the parallels to their CPU architecture do seem somewhat uncanny.
Personally I have a little mixed feelings about this having used IBM's first ever attempt at a RISC CPU in the RT6150 and got one of the first (before public release) hands on the power 1 cpu used in an RS/6000.
I also do find it interesting how an instruction set is so coveted in a day and time that compilers are so mature that to create your own ISA and adapt a compiler to accommodate it is not out of the grasp of an individual and much more easier a task than days prior. But then the real hard work is designing a CPU to run that instruction set, that becomes very much more than a one-person job. Though who knows - maybe in 10-20 years from now, the ability to design some silicon becomes as accessible as programming in SCRATCH (drag and drop style design/programming).
But back to the Power Architecture, what is IBM's real plans here - charity/PR or a way to shift processor design into a realm of outsourcing that enables them to focus upon their more profitable streams and focus - like services? Maybe.
Though I do feel they may of missed the boat upon this, what with Apple at one stage a big Power cpu user and now looking to shift towards ARM, of which they design their own CPU's. Could Apple embrace this opertunity or would the transition be a cross to hard for them to bear? That is for me one of the signs on how this will traction. If Apple shows interest, then this could be epicly big. Not saying that if they don't that it will fail. But do feel that had they did this 5-10 years ago, things would of been more likely to traction and grow. Today, beyond some niche area's, I'm not so sure it will be as epic as it could of been and may just be a case of too little too late. Equally I'm mindful that whilst IMHO it is too late to have as big an impact as it could of been, it may prove damaging towards RISC V CPU progression and by splitting the options open - may well damage both and only aid the existing Intel and Arm options out there.
https://www.phoronix.com/scan.php?page=article&item=rome-pow...
That seems to make the benchmarks entirely useless for gauging Power performance for the server market. Not saying it didn’t fall behind its competitors (I’d be surprised if it didn’t given how much more AMD and ARM makers have been investing in chip development), but these benchmarks don’t tell us that.
It may have academic value, but practically speaking, how does this matter?
Just asking the crowd in case there is a perspective I am missing.
PPC machines are also required if for some reason you're still running AIX, IIRC.
It's probably a way of making more money out of those legacy customers
(yes because if you're still running AIX you're deep down in legacy territory, and you might just want to look at recompiling your stuff for Linux but chose to suffer instead)
IBM will let you run your ancient apps, and on mainframes even apps back as far as OS/360.
But you have to pay for it.
In a good year, Apple sold about 3 million PowerPC Macs. Sony and Microsoft sold about 80 million PS3s and Xbox 360s each. I think one could make the case that those Macs were of questionable relevance to begin with... (yes I know they used G5 Macs for dev machines)