While on the hunt for "just any" working graphics adapter for my resurrected RS/6000 model 7006-42W, I got my fingers on a whole bunch of different cards.
After fiddeling, I found most of them working in my system, and some older models not, which are also not part of the "IBM feature availibility matrix", so that's probably fine.

Curiosity stuck me to find out which one would be the best card for the next vintage MCA RS/6000 Quake tournament, and I wanted to be prepared.
Fortunately, Quake for AIX has, just like the x86 version, some benchmarking built-in. It can be started from the console command line (with "+timedemo demo1" argument) or via the in-game debug console (with "timedemo demo1"), and it basically runs a defined gameplay demonstration sequence while measuring the time it takes to display all the rendered images.

So I nailed down the specs of my test system:

Model 7006-42W
128 MB RAM in 4x Kingston KTC1481
120 MHz PowerPC 604
No L2 cache

For software I'm sticking with

AIX 4.3.3 ML11 + post-ML11 fixes
Quake for AIX 1.07 (latest version) with PowerPC optimizations

Using the command line

../quake.sw.wimp (-2) -nosound -nolan -nojoy -nocdaudio +stopdemo +timedemo demo1

disables some non-relevant parts of the game and immediately starts the benchmark, which takes about 1 minute.
I did the test for the original window size and with pixel doubling (-2 parameter).
Each test was repeated 3 times, and the presented results are averaged, although they're not really far distributed for the individual runs.
CPU load was more or less precisely observed in the "top" utility.

For the original window size I measured:

1-5 POWER Gt4xi 24-bit 20.5 FPS @ 89% CPU
1-8 POWER Gt4e 20.5 FPS @ 90% CPU
1-9 POWER Gt3i 20.1 FPS @ 91% CPU
1-E POWER GXT150m 21.8 FPS @ 100% CPU
POWER GXT500D 20.0 FPS @ 100% CPU
With pixel doubling, my results are the following:
1-5 POWERGt4xi 24-bit 14.5 FPS @ 74% CPU
1-8 POWERGt4e 14.5 FPS @ 74% CPU
1-9 POWERGt3i 14.0 FPS @ 74% CPU
1-E POWERGXT150m 16.1 FPS @ 100% CPU
POWERGXT500D 13.8 FPS @ 100% CPU

What I find interesting is that the GXT500D, although it's a Class II adapter and probably the latest and most complex card in the list, AND uses the 7006's dedicated video bus instead of the MCA, doesn't perform any better or even worse than the simpler Gt3/4 family, but has the same greedy CPU load behavior like the MCA GXT150m. So this is probably the worst choice for fragging your friends.

With pixel doubling, the GPU processing power seems to be the limiting factor of the Gt3/4 adapters, while the GXTs perform slightly better.

I assume the generally high CPU load of the GXTs, compared to the Gt3/4 cards, to be driver-related, but who knows.

Personally, I would have expected a higher all-over difference across the cards. I'm not sure how much of OpenGL is supported by Quake at all; the RTE is definitely installed, but it looks like most of the rendering ist done in the CPU anyway.

So yeah, if you got cycles to spare, you'll want a GXT150m to max out your gameplay.
Sure, Quake is not the only thing you want from your RS/6000, so the picture may be different if you're benchmarking CATIA or so.

The adapters I couldn't get to work with my system are

1-1 Color Graphics Display Adapter
1-2 Grayscale Graphics Display Adapter
1-3 SGI IrisVision 8-bit

By looking at them, it becomes clear these are older types, with copyrights and IC datestamps in the late 80's, and probably not supported in this mid-90's machine.
Maybe I'll repeat the same test with my model 7013-59H once I have completed all the parts.

Any thoughts, remarks, discussion and additions are very welcome.
~Christian

Christian, based on the used launch command (quake.sw.wimp), it seems
that you are running the software renderer build of the game.

What that means is that the actual rendering is done in software on the
CPU side - including the scene transformation, lighting (most of which
is pre-baked in from of lightmaps, except for dlights), texturing, and
rasterization.

The video adapter acts as a mere frame buffer (storage for the already
finished and rasterized scene). That's why there's a very little
framerate difference between the adapters. Pretty much the only thing
that matters here is how fast is the adapter at RAM to VRAM transfers
(driver implementation, bus speed, VRAM subsystem...).

So, what you were really testing is the CPU performance, the RAM-VRAM
throughput, and then the software/driver side. The actual "accelerator"
was sitting there, doing nothing.

I'm not familiar with the later high-end RS/6000 GPUs, but I know that
the older MCA-based adapters can't do texture mapping, which renders
them useless for 3D gaming...

On 06.12.2021 14:22, Christian Holzapfel wrote:
> While on the hunt for "just any" working graphics adapter for my resurrected RS/6000 model 7006-42W, I got my fingers on a whole bunch of different cards.
> After fiddeling, I found most of them working in my system, and some older models not, which are also not part of the "IBM feature availibility matrix", so that's probably fine.
>
> Curiosity stuck me to find out which one would be the best card for the next vintage MCA RS/6000 Quake tournament, and I wanted to be prepared.
> Fortunately, Quake for AIX has, just like the x86 version, some benchmarking built-in. It can be started from the console command line (with "+timedemo demo1" argument) or via the in-game debug console (with "timedemo demo1"), and it basically runs a defined gameplay demonstration sequence while measuring the time it takes to display all the rendered images.
>
> So I nailed down the specs of my test system:
>
> Model 7006-42W
> 128 MB RAM in 4x Kingston KTC1481
> 120 MHz PowerPC 604
> No L2 cache
>
> For software I'm sticking with
>
> AIX 4.3.3 ML11 + post-ML11 fixes
> Quake for AIX 1.07 (latest version) with PowerPC optimizations
>
>
> Using the command line
>
> ./quake.sw.wimp (-2) -nosound -nolan -nojoy -nocdaudio +stopdemo +timedemo demo1
>
> disables some non-relevant parts of the game and immediately starts the benchmark, which takes about 1 minute.
> I did the test for the original window size and with pixel doubling (-2 parameter).
> Each test was repeated 3 times, and the presented results are averaged, although they're not really far distributed for the individual runs.
> CPU load was more or less precisely observed in the "top" utility.
>
> For the original window size I measured:
>
> 1-5 POWER Gt4xi 24-bit 20.5 FPS @ 89% CPU
> 1-8 POWER Gt4e 20.5 FPS @ 90% CPU
> 1-9 POWER Gt3i 20.1 FPS @ 91% CPU
> 1-E POWER GXT150m 21.8 FPS @ 100% CPU
> POWER GXT500D 20.0 FPS @ 100% CPU
>
> With pixel doubling, my results are the following:
>
> 1-5 POWERGt4xi 24-bit 14.5 FPS @ 74% CPU
> 1-8 POWERGt4e 14.5 FPS @ 74% CPU
> 1-9 POWERGt3i 14.0 FPS @ 74% CPU
> 1-E POWERGXT150m 16.1 FPS @ 100% CPU
> POWERGXT500D 13.8 FPS @ 100% CPU
>
> What I find interesting is that the GXT500D, although it's a Class II adapter and probably the latest and most complex card in the list, AND uses the 7006's dedicated video bus instead of the MCA, doesn't perform any better or even worse than the simpler Gt3/4 family, but has the same greedy CPU load behavior like the MCA GXT150m. So this is probably the worst choice for fragging your friends.
>
> With pixel doubling, the GPU processing power seems to be the limiting factor of the Gt3/4 adapters, while the GXTs perform slightly better.
>
> I assume the generally high CPU load of the GXTs, compared to the Gt3/4 cards, to be driver-related, but who knows.
>
> Personally, I would have expected a higher all-over difference across the cards. I'm not sure how much of OpenGL is supported by Quake at all; the RTE is definitely installed, but it looks like most of the rendering ist done in the CPU anyway.
>
> So yeah, if you got cycles to spare, you'll want a GXT150m to max out your gameplay.
> Sure, Quake is not the only thing you want from your RS/6000, so the picture may be different if you're benchmarking CATIA or so.
>
>
> The adapters I couldn't get to work with my system are
>
> 1-1 Color Graphics Display Adapter
> 1-2 Grayscale Graphics Display Adapter
> 1-3 SGI IrisVision 8-bit
>
> By looking at them, it becomes clear these are older types, with copyrights and IC datestamps in the late 80's, and probably not supported in this mid-90's machine.
> Maybe I'll repeat the same test with my model 7013-59H once I have completed all the parts.
>
>
> Any thoughts, remarks, discussion and additions are very welcome.
> ~Christian
>

On 12/6/21 06:22, Christian Holzapfel wrote:
> While on the hunt for "just any" working graphics adapter for my resurrected RS/6000 model 7006-42W, I got my fingers on a whole bunch of different cards.
> After fiddeling, I found most of them working in my system, and some older models not, which are also not part of the "IBM feature availibility matrix", so that's probably fine.
>
> Curiosity stuck me to find out which one would be the best card for the next vintage MCA RS/6000 Quake tournament, and I wanted to be prepared.
> Fortunately, Quake for AIX has, just like the x86 version, some benchmarking built-in. It can be started from the console command line (with "+timedemo demo1" argument) or via the in-game debug console (with "timedemo demo1"), and it basically runs a defined gameplay demonstration sequence while measuring the time it takes to display all the rendered images.
>
> So I nailed down the specs of my test system:
>
> Model 7006-42W
> 128 MB RAM in 4x Kingston KTC1481
> 120 MHz PowerPC 604
> No L2 cache
>
> For software I'm sticking with
>
> AIX 4.3.3 ML11 + post-ML11 fixes
> Quake for AIX 1.07 (latest version) with PowerPC optimizations
>
>
> Using the command line
>
> ./quake.sw.wimp (-2) -nosound -nolan -nojoy -nocdaudio +stopdemo +timedemo demo1
>
> disables some non-relevant parts of the game and immediately starts the benchmark, which takes about 1 minute.
> I did the test for the original window size and with pixel doubling (-2 parameter).
> Each test was repeated 3 times, and the presented results are averaged, although they're not really far distributed for the individual runs.
> CPU load was more or less precisely observed in the "top" utility.
>
> For the original window size I measured:
>
> 1-5 POWER Gt4xi 24-bit 20.5 FPS @ 89% CPU
> 1-8 POWER Gt4e 20.5 FPS @ 90% CPU
> 1-9 POWER Gt3i 20.1 FPS @ 91% CPU
> 1-E POWER GXT150m 21.8 FPS @ 100% CPU
> POWER GXT500D 20.0 FPS @ 100% CPU
>
> With pixel doubling, my results are the following:
>
> 1-5 POWERGt4xi 24-bit 14.5 FPS @ 74% CPU
> 1-8 POWERGt4e 14.5 FPS @ 74% CPU
> 1-9 POWERGt3i 14.0 FPS @ 74% CPU
> 1-E POWERGXT150m 16.1 FPS @ 100% CPU
> POWERGXT500D 13.8 FPS @ 100% CPU
>
> What I find interesting is that the GXT500D, although it's a Class II adapter and probably the latest and most complex card in the list, AND uses the 7006's dedicated video bus instead of the MCA, doesn't perform any better or even worse than the simpler Gt3/4 family, but has the same greedy CPU load behavior like the MCA GXT150m. So this is probably the worst choice for fragging your friends.
>

It's seems indicative of a software bottleneck, likely the interaction
between Xwindows and the rendering mode. In other words, all of these
adapters are getting stuck in the same place and are not the limiter?

> With pixel doubling, the GPU processing power seems to be the limiting factor of the Gt3/4 adapters, while the GXTs perform slightly better.
>
> I assume the generally high CPU load of the GXTs, compared to the Gt3/4 cards, to be driver-related, but who knows.
>
> Personally, I would have expected a higher all-over difference across the cards. I'm not sure how much of OpenGL is supported by Quake at all; the RTE is definitely installed, but it looks like most of the rendering ist done in the CPU anyway.

I don't think it is being used at all w/quake, this is CPU rendering and
blitting without direct rendering.

> So yeah, if you got cycles to spare, you'll want a GXT150m to max out your gameplay.
> Sure, Quake is not the only thing you want from your RS/6000, so the picture may be different if you're benchmarking CATIA or so.
>
>
> The adapters I couldn't get to work with my system are
>
> 1-1 Color Graphics Display Adapter
> 1-2 Grayscale Graphics Display Adapter
> 1-3 SGI IrisVision 8-bit
>
> By looking at them, it becomes clear these are older types, with copyrights and IC datestamps in the late 80's, and probably not supported in this mid-90's machine.
> Maybe I'll repeat the same test with my model 7013-59H once I have completed all the parts.
>
>
> Any thoughts, remarks, discussion and additions are very welcome.
> ~Christian
>

So. In a small, quiet voice... have you delved into using the AIX ODM [I
think...] to create ADFs for RS/6000 adapters being used in PS/2s?

https://ardent-tool.com/RS6000/RS6000_4-4.html

On 12/6/2021 07:22, Christian Holzapfel wrote:
> While on the hunt for "just any" working graphics adapter for my resurrected RS/6000 model 7006-42W, I got my fingers on a whole bunch of different cards.
> After fiddeling, I found most of them working in my system, and some older models not, which are also not part of the "IBM feature availibility matrix", so that's probably fine.
>
> Curiosity stuck me to find out which one would be the best card for the next vintage MCA RS/6000 Quake tournament, and I wanted to be prepared.
> Fortunately, Quake for AIX has, just like the x86 version, some benchmarking built-in. It can be started from the console command line (with "+timedemo demo1" argument) or via the in-game debug console (with "timedemo demo1"), and it basically runs a defined gameplay demonstration sequence while measuring the time it takes to display all the rendered images.
>
> So I nailed down the specs of my test system:
>
> Model 7006-42W
> 128 MB RAM in 4x Kingston KTC1481
> 120 MHz PowerPC 604
> No L2 cache
>
> For software I'm sticking with
>
> AIX 4.3.3 ML11 + post-ML11 fixes
> Quake for AIX 1.07 (latest version) with PowerPC optimizations
>
>
> Using the command line
>
> ./quake.sw.wimp (-2) -nosound -nolan -nojoy -nocdaudio +stopdemo +timedemo demo1
>
> disables some non-relevant parts of the game and immediately starts the benchmark, which takes about 1 minute.
> I did the test for the original window size and with pixel doubling (-2 parameter).
> Each test was repeated 3 times, and the presented results are averaged, although they're not really far distributed for the individual runs.
> CPU load was more or less precisely observed in the "top" utility.
>
> For the original window size I measured:
>
> 1-5 POWER Gt4xi 24-bit 20.5 FPS @ 89% CPU
> 1-8 POWER Gt4e 20.5 FPS @ 90% CPU
> 1-9 POWER Gt3i 20.1 FPS @ 91% CPU
> 1-E POWER GXT150m 21.8 FPS @ 100% CPU
> POWER GXT500D 20.0 FPS @ 100% CPU
>
> With pixel doubling, my results are the following:
>
> 1-5 POWERGt4xi 24-bit 14.5 FPS @ 74% CPU
> 1-8 POWERGt4e 14.5 FPS @ 74% CPU
> 1-9 POWERGt3i 14.0 FPS @ 74% CPU
> 1-E POWERGXT150m 16.1 FPS @ 100% CPU
> POWERGXT500D 13.8 FPS @ 100% CPU
>
> What I find interesting is that the GXT500D, although it's a Class II adapter and probably the latest and most complex card in the list, AND uses the 7006's dedicated video bus instead of the MCA, doesn't perform any better or even worse than the simpler Gt3/4 family, but has the same greedy CPU load behavior like the MCA GXT150m. So this is probably the worst choice for fragging your friends.
>
> With pixel doubling, the GPU processing power seems to be the limiting factor of the Gt3/4 adapters, while the GXTs perform slightly better.
>
> I assume the generally high CPU load of the GXTs, compared to the Gt3/4 cards, to be driver-related, but who knows.
>
> Personally, I would have expected a higher all-over difference across the cards. I'm not sure how much of OpenGL is supported by Quake at all; the RTE is definitely installed, but it looks like most of the rendering ist done in the CPU anyway.
>
> So yeah, if you got cycles to spare, you'll want a GXT150m to max out your gameplay.
> Sure, Quake is not the only thing you want from your RS/6000, so the picture may be different if you're benchmarking CATIA or so.
>
>
> The adapters I couldn't get to work with my system are
>
> 1-1 Color Graphics Display Adapter
> 1-2 Grayscale Graphics Display Adapter
> 1-3 SGI IrisVision 8-bit
>
> By looking at them, it becomes clear these are older types, with copyrights and IC datestamps in the late 80's, and probably not supported in this mid-90's machine.
> Maybe I'll repeat the same test with my model 7013-59H once I have completed all the parts.
>
>
> Any thoughts, remarks, discussion and additions are very welcome.
> ~Christian
>

You said it, guys. This is software rendering only.
I didn't expect the AIX Quake implementation supporting all various kinds of adapter specific 3D handling...

Louis Ohland schrieb am Montag, 6. Dezember 2021 um 17:54:59 UTC+1:
> So. In a small, quiet voice... have you delved into using the AIX ODM [I
> think...] to create ADFs for RS/6000 adapters being used in PS/2s?

Well, interesting. I'll have a look at that!