FreeBSD considering end of ppc64 support
I asked Warner Losh for clarification but he hasn't gotten back to me yet, so it's unclear if this is meant to apply to just the BE port or both. Unlike the overall industry mood generally, especially from you underannuated whippersnappers who deal with the doublethink of big-endian math in your head but little-endian code in your repos, the BSDs have historically been friendly to big-endian hardware as a means of ensuring code correctness. Indeed, his note acknowledges this. One sticking point seems to be Power ISA's own unique architectural features but I suspect it is due at least as much to the higher cost and poorer availability of the hardware, which has admittedly been thin on the ground lately.
My preference in server operating systems is a BSD-variant and FreeBSD was going to be my own migration plan when I retire my POWER6 running AIX (big-endian, natch), but this news has made me consider OpenBSD/powerpc64 instead. This port is a little suboptimal in that it states it does not presently run under a hypervisor, but it is currently well-supported and specifically tested on Raptor hardware, which is going to be the majority of the ecosystem. Of course, OpenBSD is also a famously opinionated operating system, so there may be a little bit more of a learning curve for me personally, whereas I already have experience with FreeBSD on shared x86_64 hosting and my other BSD systems all run NetBSD (on /mac68k, /cobalt, /macppc, /alpha, /hpcmips and /hpcsh). Another possibility is to try to maintain security patches and fixes on the remnant of FreeBSD 15, though I have such limited time these days I'd prefer not to pour it into another endeavour with an obvious future dead end. I'd dearly love a NetBSD option, but there isn't a NetBSD port that runs on POWER9.
I do think it's a little unfair to cite the user base as being less than "x86 or armv7" when these were extremely common architectures in their day, and a lot of people are probably still running FreeBSD on older examples to get the most money they can out of their hardware. I can't imagine their RISC-V numbers are very high either but no one is saying that should go away, and Power ISA is still very relevant in today's big-iron shops. If and until "low-end" Power11 desktop hardware is available, we'll want all the operating system support we can get too.
When Linus Torvalds had his comments the other week in regards to a kernel patch to add support for big endian to RISC-V, I think it kinda broke me a bit. I hopped into the Phoronix article about it, looked around at the comments, and talked to some people in the RISC-V Fedora Matrix room.
ReplyDeleteI was greeted by people citing other articles about how big endian is an inferior outdated thing, and their evidence was that their favorite AI chatbot at the time told them so. A bunch of, from what I could tell, non-existent studies, and regurgitated nonsense from x86 apologists, said that little endian is the only thing that makes sense, because it's the only thing that's out there. The early internet was built predominantly on 68k, and that was only big endian. The routers and infrastructure were built on 68k, PPC, and MIPS (predominantly the big endian ones for MIPS network gear). Most ARM64 processors can do big or little endian, and the software selection, and sometimes firmware selections, limit them into little endian. When we stepped past 1C1T for these devices, companies started moving to ARM, because x64 was still stupid for it, but they could get newer chips and context switching by using multiple cores (and higher clock speeds) meant they could be sloppier with their code, use some byte-swapping extensions, and just brute-force it.
Torvalds was infuriated that someone would make a patch for Linux that would enable a new generation of big endian RISC-V chips on it, because the hardware doesn't exist, and there's an extension to flip endianess of data so it can talk to big endian things without being big endian itself. If we use Linux as the test harnass, and the running system for chips we're making or attempting to make, then we're now subject to the other side of the android ecosystem, where a fork of the kernel for an android version, then a fork of that for the processor version, then a fork of that for a device manufacturers version, and nothing ever gets back to upstream and the device becomes orphaned. Mix that with the way drivers and platforms are getting removed from Linux, and maybe Linux isn't the best way forward or backward anymore. It's just going to be the sweet spot for something to support a bunch of corporate hardware, while it slowly obsoletes previously supported platforms. Maybe we need something smaller to do that sort of thing with.
BTW, my anguish with this, is that when the code stops being used and tested on big endian, when it stops being used with page sizes other than 4k, when it stops being used without 64-bit or 32-bit, or other bitness-clean code, is when that code picks up x86-isms and stops being able to be built or run on those other platforms. The people writing that code, without attempting to use it on other platforms, always get some cruft in it that they don't realize is a problem.
The lack of modern developer friendly Power hardware is significant. IBM has forgotten how scalability from small to large to larger is what allowed System/360 to dominate the computing market. At this point they tread down the same path explored earlier by Unisys of monetisating valuable legacy customers without attracting new customers to the point of irrelevancy.
ReplyDeleteOne thing is clear from Linux, the BSDs, GNU and other open source: If there is no cost-effective well-performing hardware nor any hope for such hardware in the future, the independent developers who make, test, polish and maintain open source are not interested.
Unfortunately, Raptor's systems are also too expensive to empower developers.
While corporate-sponsored open source is important, neither IBM nor can any other company on their own improve and debug the thousands of projects that form the foundations of the current software ecosystem.
Said another way, a lack of developer-friendly hardware has led to lack of interest and the bitrot reached a point where Power is not maintainable.
+Main problems:
ReplyDelete1) IBM does not care about anything outside their own business at all. They are not interested in desiging or releasing CPUs for SOHO. They have superior technology and don'T care to make it publically available. Just stupid.
2) There is almost no other company interested in PPC. A few tiny companies with specialized boards for industry but not general purpose usage and almost no one can buy them.
3) Raptor developed good products but they are much too small, too slow, their products come too late in too small batches and they are way, way too expensive. Which other hardware outside of PPC has become 3-4 times as expensive as it was on launch? None. Everything else has progressed in the last 6-7 years (3? 4 more generations of Ryzen for example with always more performance at lower and lower price). So Raptor simply cannot sell their (quite GOOD!) hardware to anybody. Only a hand ful of lovers buy it (look in to their forum, there is maybe 2 dozen people and almost no communication).
4) This whole PowerPC consortium is just bloated theory. They create fantasy papers, have people on board that talk about synthetic products but nothing ever comes out of it. SOHO boards? Graphics card? The micro board for my home lab? For universities, students at home to work with? Enthusiasts? Just nothing that anybody of us could buy.
5) Endianness: While it is theoretically good that they support both endians, developers have become even more lazy and say "okay, if they also speak little endian than I don't have to care avbout big endian. Why should I use my brain to develop robust algorthims and test their correctness if I can make it just work on one single scenario". And this is exactly what you see from most of the comments on the FreeBSD mailng list. It's like saying "hey we have cars that run on fossiles, why should we care about electric cars?".
6) Competition: The other processors are catching up in performance and they are muuuuuuch cheaper and have a much better ecosystem because they are widespread in so many industries and functionalities and different (small, cheap) boards, SDKs etc: ARM, RISC-V etc. Everyone can buy them, test them, bring their apps to it. ARM, Ampere, NVIDIA etc. are all growing in server market. One day when IBM raises too high prices for their stuff the customers will jump off and buy ARM or NVIDIA or even RISC-V instead. Then the market for IBM will quickly break down and they will loose many customers. So big and slow as IBM is, it would take them 5-10 years to come up with something that is still competitive. But customers you lost once won't come back.
Regarding the micro board, you can help funding this: https://www.powerpc-notebook.org/2025/12/pcb-design-costs-and-timing/
DeleteThis PPC notebook/desktop/notebook is as real and useful as the micro boards. It's just hot air, nothing usable. IT's much more than funding what they need. Look at their history, they got 68,000 EUR for the past stages over more than 10 years: https://www.powerpc-notebook.org/list-of-donors/. Plus more than 3,000 EUR now for the current stage. 71,000 EUR in 10 years, all stages financed by donors, and they still don't have it ready!
DeleteI'd love to have a modern PPC notebook but they are working on it for 10 years, they failed with the notebook so they stepped back to design a desktop mainboard and now want to change over again to the notebook board because they think now they have the skills and software and everything right. I'D LOVE to buy a new PPC notebook but I'm sceptical when I look at their history. If it takes them another year AT LEAST (and another and another ... which is how it went for the past 10 years, they sooo often said "yeah, just a few more months then we can sell it", "end of the yerar the board will be ready" ...) This never came true. I'd rather invest in the Mirari crew : https://mirari.vitasys.nl/the-first-rebirth/ . What they are delivering in just 1 and a half year is fantasic! Linux runs OOTB, MorphOS runs since summer (I read that it took only one dev and 6 days to do the main port, now they are developing the remaining drivers). It is a desktop, no notebook. But this shows what a good team can achieve. Now image IBM or Raptor funding these guys for production of a few thousand boards!
I'd like to invite to ArchPOWER discord, has more than +150 ppl and counting
ReplyDeletehttps://discord.gg/HntKjSTrVe
https://archlinuxpower.org/
It supports many powerpc platforms
Regards,
Link.
Why drag people to yet another closed source centralised chat platform that can be shut down by the owner at any time? Use open platforms instead. That Microsoft is all in MS Teams and such, I understand. But why open source or open hardware goes behind the Discord fence I don't understand. It doesn't make sense.
DeleteYou can see, there was a lot of push back on retiring powerpc64*, and even some renewed interest in expanding it, so it won't be gone in 16. I'll personally continue running FreeBSD/powerpc* on all my powerpc machines, even maintaining 32-bit FreeBSD/powerpc (and powerpcspe) while I feel so inclined.
ReplyDeleteTHis soudns good! So not only the LE variant is continued but also BE?
DeleteBE will likely continue for 16. The overhead of maintaining BE vs LE is pretty small, especially on this architecture. The rest is all drivers and ports, and drivers are pretty easy to keep up to date with endianness for the most part. Ports may be problematic, but PowerPC is a Tier 2 platform anyway, so ports support is "best effort", and we do have a couple dedicated PPC ports maintainers who do a fantastic job keeping them at least building.
Delete