IBM's POWER9 retrospective is a little one-sided


I'm not going to fault IBM taking a victory lap with the POWER9 in their one-year anniversary retrospective. It's a kick-ass processor; that's why I'm typing this on one. I'm not even going to fault them for biasing it towards their own server products, because that's what IBM sells and they're a business and they want to sell their own stuff. And IBM maintaining financial health gives them the R&D capacity to make the POWER10 even more awesome, so bring on the salesdroids.

But while Google got a shout-out with their bespoke POWER9 Zaius server platform, now in production, IBM seems to have forgotten that Power ISA is making a triumphant return to the desktop in a form that's gotten at least as much press as Zaius/Barreleye, probably sold more units, and is actually in the hands of real end users who are using it as their real computers right now. Hmm, I wonder what computer that could be?

Let's consider the historical perspective: the last major Power desktop system that didn't come from IBM was of course the Power Mac G5 Quad, which was replaced by the Mac Pro in 2006. (The last Power workstations from IBM were the IntelliStation POWER 185 and 285. This pair of machines outlasted the Quad G5 until 2009 but they preferentially ran AIX, and weren't available in large numbers.) While PowerPC chips owned the game console market for a period of time (Wii/Wii U, PlayStation 3 and Xbox 360), Apple's Intel transition meant the only remaining third-party PowerPC desktops were the AmigaOne machines. I like Amigas fine, but these systems so far have been rather underpowered due to their use of embedded designs and fairly expensive due to the small market, boutique production runs and distinctly poor economies of scale. Frankly, depending on how big a detractor you are, PowerPC hadn't been competitive against x86 on the desktop since the early G4 days, and no one other than IBM had shipped a top-tier Power ISA desktop in over 12 years.

Now we have not only a Power ISA CPU that is performance-competitive with current x86_64 offerings, but an entire third-party libre workstation built around them that you can order and get shipped to your house right now. The cost of a full Talos II only seems steep until you consider how much a Xeon box in the same ballpark will run you, and the delta seems much more reasonable then. If Blackbird is successful at establishing a "low end" POWER9 machine with a more amenable price, we could see Power ISA start to become a major desktop player once again, especially as the software support situation continues to improve by leaps and bounds and people realize what a liability blackboxes like Intel Management Engine are. Even if you're not in the market for a T2 right now, you'll be more likely to have a real choice in workstations when you do. And that's good news for everybody.

Seems like IBM could have mentioned that.

Comments

  1. You have completely forgotten YDL PowerStation! http://linuxdevices.org/quad-core-powerpc-workstation-runs-linux/ --- read the date of publication. 2008! :-)

    ReplyDelete
    Replies
    1. Did those ship? I only ever saw a couple and I thought those were prototypes.

      Delete
    2. It seems:
      http://www.amibay.com/showthread.php?36477-ITALY-YDL-Powerstation-dual-2-5GHz-970MP-(quad-G5)

      Delete
    3. Was about to bring YellowDogLinux PowerStation up myself. :) I tried finding one, but no success. They could have up to 32 GB of RAM, as opposed to Quad's 16 GB, but since it couldn't boot Mac OS X for software reasons, it was more of a whim to me.

      The recent Amiga boxes are kinda neat, but IIRC basically obsolete compared to the Quad G5 in terms of hardware.

      I still want to see Mac OS 9 boot on a Quad (or Talos II... as if!). Even though what doing so entails requires an insane (!) amount of time, knowledge & dedication. Well, at least we've got the Mac mini G4 booting OS 9 natively, so cheers I guess. :)

      Delete
    4. What about SheepShaver, since it's OpenSource.
      https://de.wikipedia.org/wiki/SheepShaver

      Delete
    5. You mean for having Mac OS 9 in G5s and modern PowerPCs? Definitely something nice to have, but still quite behind in terms of software stability and other technical problems (games don't run as smoothly, for instance. You can easily verify that by playing the first Escape Velocity on native OS 9 vs. SheepShaver or even Basilisk II).

      Emulation & virtualization are very nice, practical & useful to have, but almost never a true replacement for the real thing. Although sticking to them is by far more attractive than deciding to keep writing drivers for new hardware, especially considering OS 9 isn't open-source for us to look inside it better...

      QEMU not so long ago managed to emulate the latest OS 9, OS 9.2.2, although without sound & a few other issues, but I'm happy to see that is at least getting somewhere. Now, I only wish I could compile HQEMU under OS X (rather than GNU/Linux) on a G5 Quad...

      Delete
  2. To be honest, the Quad G5 was competitive well into 2006, as you can see here:
    https://web.archive.org/web/20060901063318/https://setiathome.berkeley.edu/top_hosts.php

    I am on rank two 8-)
    I don't know what happened to the "optimized" worker today. But with the POWER9 something similar should be possible.

    SETI/ Boinc is OpenSource, but I have no idea where the source code is.

    I still remember a statement from "hobold" on SIMDtech.org, shortly after the switch:
    "I had a conversation with an engineer from a certain fruit company today and he said that even Intel engineers were having problems getting SSE2/3 versions of some of the Apple Altivec sample code running at anything better than half the speed of the Altivec code, and this on a CPU with twice the clock speed of a G5. Steve can sit in his distortion field all he wants but that doesn't change the fact that Altivec is far superior to SSE2/3. In fact, in many other ways the G5 and PPC architecture is superior to Intel and x86 (but not in all ways of course). Many of us here know this through experience. The way I see it this is for mobile chips and that's it. That's the only weak point of the current PPC architecture and currently the strong point for Intel. AMD is beating them on the Desktop and Intel is even having to adopt AMD's 64-bit architecture."

    ReplyDelete
    Replies
    1. Do you have any source for that segment? That's a lot of conversational text for a single memory.

      It would be something else to have written proof from an Apple engineer admitting that PowerPC (and likely by extension, POWER) was very much superior to Intel besides portable usage.

      I'm certain the landscape today has not changed at all.

      Delete
    2. Sorry, I don’t have a screenshot from the post. The statement was made on the „AltiVec Forum“ from Freescale. I C&P it at that time. 

Here is another link:
https://arstechnica.com/civis/viewtopic.php?p=8206227&sid=0d309a98f71e600404ea54f35e343eb1#p8206227

      So, at the end, I could’t prove it.

      Delete
    3. While I do think OpenPOWER is great and I'll probably end up buying one as a daily driver in the future (I'm tight on budget and my current workstation is only a couple years old), your statement is not much true. It was at the time SSE2 and SSSE3 came out, but nowadays most resource-intensive program use AVX or at the very least SSE4 and backporting those optimizations to PowerISA isn't that straightforward as you seem to imply (if an ISA were vastly superior to another, porting the former optimizations to the latter wouldn't be particularly difficult).. Then again the POWER architecture is much "stronger" than x86 from both a scalability and reliability point of view, that's 100% true, but x86 is far more mature (captain obvious here duh^^).

      Delete
    4. Source: http://openpowerfoundation.org/wp-content/uploads/resources/Vector-Intrinsics/Vector-Intrinsics-20180306.pdf#%5B%7B%22num%22%3A52%2C%22gen%22%3A0%7D%2C%7B%22name%22%3A%22XYZ%22%7D%2C72%2C223.096%2Cnull%5D

      Delete
    5. "I had a conversation with an engineer from a certain fruit company today and he said that even Intel engineers were having problems getting SSE2/3 versions of some of the Apple Altivec sample code"
      If it is true it was likely referring to Pentium 4 or Pentium M/the first 32bit Core Duo, which had particularly weak SSE unit (for example x264 detected the chip and only used MMX because that was faster despite just 64bit width).
      I doubt it remained true when Core 2 came out in 2006, made the SSE unit fully 128bit (SIMD ops were suddenly 2x faster) and added SSSE3.

      BTW as Mad_Hatter says... today x86 has 256bit SIMD width and 512bit is coming. IIRC Power has stayed on 128bit so for SIMD/multimedia code it is arguably behind (though it has lots of brute force...).

      Delete
  3. There is more hardware to mention, namely the Pegasos family of German company bPlan: http://www.bplan-gmbh.de/home-en.html (no, not Genesi) including Pegasos (I), Pegasos II, EFIKA, and ODW - the first open source based PowerPC computer: https://en.wikipedia.org/wiki/Pegasos

    ReplyDelete

Post a Comment

Comments are subject to moderation. Be nice.