viva PowerPC

Thoughts and questions on Cameron's Blackbird semi-review

Just yesterday (at least from my timezone point of view) Cameron published A semi-review of the Raptor Blackbird: POWER9 on the cheap(er). Go and read it, before you continue reading my text. Go. Do it.

I read it twice and wasn't happy from what I found there. At first I wanted to write my response on PNUT, then I thought I'd send him an e-mail, but then I said: hey, there may be another people like me, so why not doing it in public? So...

I used PowerPC-based computers as my primary between 2004 and 2018. Couple of years before that I had a Mac on my desk as a secondary machine as I do now, in the last two years. I'm more Linux/Unix kind of guy than a Mac one, so even though I like Mac OS 9, it was much more convenient for me to use Linux on a generic PC box as a primary before Mac OS X matured enough to replace it for my daily usage. That happened somewhere between 10.3 and 10.4, then in 2005 i bought a new iBook G4, after that a second-hand PowerBook G4, PowerMac G4, PowerMac G5 and that's pretty much is where I left the platform in the beginning of 2018 as you can read in one of my earlier posts.

When I decided to retire my G5, there was no POWER/PowerPC based replacement available. Talos II was and still is out of question, I'm simply not in the position to buy a computer with the price tag of a cheap new car, no matter how open, fast or POWER is it and no matter how much I'd like to have it. But Blackbird is completely different kind of bird - the board, a quad-core POWER9 and a heatsink with oversea shipping and local taxes will be about $1850, which is still way above any mainstream computer with the same performance, but it's doable. There are several local shops with 16 GB DDR4 ECC Reg. modules around $140, I already have some SATA SSD and for sure will dig out an ATX case somewhere. It won't be exactly a nice, eye-catching computer, but it will be my own POWER9 machine and nothing else matters. I'm trying to get back in the good old PowerPC mood and it seems that I'm already on the right track, so the last thing I want to read about my possible (and probable) future computer is, that it's not even fit to boot the OS properly to working desktop and when it boots it's not exactly fast enough to do things I do on 5-year-old Core-i5 laptop or even a G4 Mac Mini.

So, Cameron, here are my questions, I don't think there is anyone more eligible to answer them:

  1. Is there any possibility that you are spoiled by the Talos? Would you think that quad-core Blackbird is not fast enough as a primary daily machine if you switched to it from your quad-core PowerMac, not Talos?
  2. How is Mac OS X virtualization performing? My G4 Mini has 812 points in Geekbench2, my PowerMac G5 had 1745 points in Geekbench2 and the absolutely fastest PowerPC machine I ever had - RackMac3,1 - was just above 2000 points. Every single POWER9 core should be faster than any of this.
  3. How does the machine perform with some light window manager? I use things like i3 even on Xeon-based workstations, because I don't think the looks of the desktop is worth any performance penalty, even if you have plenty of performance.
  4. How are supposedly 2D tasks like batch photo editing performing without dedicated GPU? These days there is GPU code everywhere, so what's the penalty here?

I think that at least some of this could be of interest for more people than just me - the group of PowerPC enthusiasts eager to replace their PowerMac with something newer, may be as large as people wanting POWER9 HTPC or a low-end POWER9 machine without any previous experience with the architecture.

Your comments: 7

[1] ClassicHasClass (June 9, 2019 at 10:19 PM CEST)
I'll bite. These are strictly my opinions only! I'll also reiterate that the review's intention was to be super duper cheap and the system was specced with that in mind, as a warning for people not to go too low and then be unhappy with the result. That said:

1. Am I spoiled by my dual-4 Talos II? You bet, I think it's great (this reply is being typed on it). But don't think I couldn't make a single-4 work, and in basic use it's far faster than the G5 — just try even this modest configuration in Firefox even without a JIT, it runs rings around TenFourFox on virtually every benchmark. However, even that G5 will have a GPU. Remember, I intentionally tried to buy as low as I could.

If you want to optimize for one thing than the other, I think I feel the lack of the GPU more because a lot of stuff depends on it as you correctly point out. But I still stand by the 8-core recommendation just because of how much is running in the background on mainstream Linux desktop distros, though see my answer to question 3.

2. Virtualization performance on the Blackbird is the same as on the T2 since at that level they're the same design. Both are somewhat hobbled because of dcbz32 emulation, plus KVM-PR must emulate supervisor instructions, but even with all the problem state overhead it was still faster than the Quad at least in Reduced mode. What really needs to happen is to get QEMU to the point where it emulates G5 hardware, which is a better match for the POWER9 (dcbz128, no mcrxr, etc.). Right now it looks like a G4 and Mac OS X treats it that way. I have hacks for KVM-PR to get around this by monkeypatching the commpage and that improves Geekbench by nearly 200 points but that doesn't fix the problem for user applications. Also note there is no GPU, of course. See the respective Talospace articles on that, starting with

You may be surprised to hear this but I actually use TCG CPU emulation for installations and other I/O heavy operations that run a lot of supervisor-level code because it involves less kernel-user context switching and the throughput is generally more reliable. Once the OS and apps are installed, though, then I run in KVM for better CPU performance.

3. I intentionally used Fedora GNOME so that I had as direct a comparison between my T2 and the Blackbird, as sort of a worst case. (Again: don't run Wayland, run Xorg.) The Blackbird seems similarly responsive in GNOME's UI as the T2 other than some UI animations not being as smooth. Otherwise, basic functionality felt nearly the same. It's not that I think Xfce or i3 some other light window manager would run badly — clearly they would run well, if GNOME runs well; it's more that I don't think GNOME is the rate limiting step here. But if those require fewer threads you might get a little extra processing headroom for other apps.

4. I don't know, I didn't try anything like that. I ran LibreOffice and a few other things, and those worked fine, but I didn't do any big/multiple image work or benchmarks that simulated it. A lot more things have AltiVec in them than they used to, though, so this may be less of a problem than it used to be.
[2] ClassicHasClass (June 9, 2019 at 10:34 PM CEST)
Also, while I think of it, let me respond directly to this statement: "it's not even fit to boot the OS properly to working desktop and when it boots it's not exactly fast enough to do things I do on 5-year-old Core-i5 laptop or even a G4 Mac Mini."

If I gave that impression, I apologize. The horror show when I first brought it up was Wayland-related. Switching to Xorg, it was no longer an issue. As far as 1080p playback, that does seem to be app-related, since VLC was really the one app that exhibited the most problems. Firefox and mplayer did fine and most other stuff worked too as long as you weren't doing something GPU dependent.

Both of those systems are going to have more on board than this "stripper" configuration did anyway. The G4 mini has a 9200 and this system has a 2D framebuffer, and LLVMpipe is a poor substitute for hardware. It's just the way it is. I think this machine outclasses both those systems, for what it's worth.
[3] Jan Olšan (June 9, 2019 at 10:57 PM CEST)
I commented about this under the original blogpost but I believe the whole problem was the use of the ASpeed thing instead of a GPU. Those chips are just not meant for such use. The result would likely be much different if some baseline cheap Radeon was used, with proper X11 accelerated stack.
I don't agree with 8core being more needed than the GPU. Background tasks don't take that much CPU time. Or should not. (I have a cheap qualdcore x86 on Windows 10 and there is never any issue, i'd say that thing doesn't even require one "reserve" core).
And anyway, you can almost get a suitable card for free/in some dumpster, so it doesn't make sense to compromise here. AFAIK open source driver for Radeons works on Talos/Power9 so if you check for some guaranteed-working type, you should be fine, no need to shell for expensive Radeon Pros.

Of course, it's not a good idea to aim for a Power9 HTPC given the lack of SIMD code in ffmpeg and video players, but that is probably manageable, the CPU should still have substantial processing power. You can watch full HD video on things like Core 2 Duo. Power9 will cause high CPU usage, but barring bugs and such stuff, it should still work. And in other tasks, it should really work okay for general use. If 8 years old i7-2600K can be a luxurious desktop machine (and a dualcore would likely be too), then this CPU should too, unless Power9 is seriously bad which people insist it's not.
[4] Anonymous (June 9, 2019 at 11:29 PM CEST)
"Of course, it's not a good idea to aim for a Power9 HTPC given the lack of SIMD code in ffmpeg and video players, but that is probably manageable, the CPU should still have substantial processing power"

Quick note on that — I personally run live HD transcode into x264 / AAC / HLS on a single 4-core box (game streaming / capture). CPU use didn't even come close to pegging two threads, let alone an entire core — in fact the CPU is mostly idle in top, and it feels like a waste of the 4-core chip in that application. I also run MythTV on a small Talos II build with a Polaris GPU with OTA broadcast in 1080p and my 1080p 60Hz x264 capture files — it doesn't even break a sweat.
[5] Jan Olšan (June 17, 2019 at 3:52 PM CEST)

Teda nevím, jestli ty testy jsou úplně stejný (ta kompilace Linuxu a GCC mi ale přijde jako že je), ale kupodivu to 4c/16t Power9 dostává docela drsně čočku od 6c/6t Intelu.

To jsem teda zvědavej, jak ten Blackbird dopadne v klasický recenzi (tu ještě Phoronix nevydal, nebo jsem ji přehlídl?). J eto divný, protože z těch recenzí velkýho Talosu se nezdálo, že by jádro Power9 bylo takhle slabý.
[6] Logout (June 25, 2019 at 11:29 AM CEST)
Vůbec si netroufám říct, jestli jsou ty časy srovnatelný, když předpokládám že jak GCC 8.2 tak ten kernel je na každý architektuře překládán pro nativní a tudíž jde o zpracování jiného kódu.
[7] Jan Olšan (June 27, 2019 at 8:34 PM CEST)
A jo, to je vlastně pravda. Asi to na x86 z praktických důvodů optimalizovali na na rychlost, kdežto na PowerPC se vykašlali.

Já se snažil najít nějaký testy, kde by se dalo srovnávat (takže v tom je to GCC fail), protože Phoronix pořád ne a ne dodělat test výkonu, ale nic jinýho jsem tam zběžně neviděl.

Taky by se do toho mohli pustit, když už to železo mají.

Leave a comment

Please enclose every link in your comment in a pair of square brackets, ie. []. Thank you.

Your comment
Type in digits from the string on the right *8*2*7*3*6*