VOGONS


Exploring the performance of the Matrox G550

Topic actions

Reply 100 of 119, by vvbee

User metadata
Rank Oldbie
Rank
Oldbie

The question of hardware T&L is resurrected. The answer is, the G550 does appear to have functional hardware T&L for DX7 and DX8 (not OpenGL), starting sometime around driver 6.7x. It can be exposed in 6.83 by changing g550dd32.dll's bytes 0xeac1 (DX7) and 0x59f11 (DX8) from 74 to eb.

The attachment g550hwtnl.gif is no longer available

Based on brief testing, it appears to work generally speaking, but some types of polygons have issues. Game 1 in 3DMark2000 for example has glitched terrain while game 2 renders with few to no glitches. On a P4 2.4 GHz test machine hardware T&L loses a bunch of performance, around 50% in 3DMark2000. In Evolva hardware T&L gets around 27 FPS while software T&L hits the vsync limit at 60. I don't know whether this T&L would be of use on a slower CPU, maybe.

Even though hardware T&L isn't exposed by default, apparently fog is always computed by the on-board T&L engine, unlike on the G400 which doesn't appear to have any. Says Claude's analysis.

Reply 101 of 119, by The Serpent Rider

User metadata
Rank l33t++
Rank
l33t++

apparently fog is always computed by the on-board T&L engine

Doesn't seem to be right, unless applied to OpenGL.

I must be some kind of standard: the anonymous gangbanger of the 21st century.

Reply 102 of 119, by vvbee

User metadata
Rank Oldbie
Rank
Oldbie

It seems very right since that's what the driver appears to get up to. If you can show otherwise then please do.

Reply 103 of 119, by The Serpent Rider

User metadata
Rank l33t++
Rank
l33t++

It straight up ignores software T&L D3D option?

I must be some kind of standard: the anonymous gangbanger of the 21st century.

Reply 104 of 119, by vvbee

User metadata
Rank Oldbie
Rank
Oldbie

There's no T&L pre-DX7 so yes. But also there's no specific DX option to compute fog by the driver on the CPU that I'm aware of. I'm talking about table fog by the way.

Reply 105 of 119, by dr.zeissler

User metadata
Rank l33t
Rank
l33t

If you are comparing G400 to G450 / G550 these things came to my mind.

- Older Games benefit more from the 128Bit memory interface of the G400 than the DDR-Ram of the later cards (450/550) with 64Bit interface
- G450 does support EMBM in Expendable by RAGE, G550 does not work.
- G450/550 both lack VSYNC in OpenGL mode that is essentiel for lot's of the remakes of classic intro's and cracktros of the 16Bit area (Atari/Amiga)
- G450 has a working Win3x driver for WinOS/2 but no native Win3x driver itself. Don't know if the G550 works too. G400 should have a native Win3x driver.

The second texture-unit of the G550 does not benefit that much in real performance. In short the G550 is mostly only 20-30% faster, but perhaps this depends on the game (engine).

Doc

Last edited by dr.zeissler on 2026-11-17, 17:31. Edited 1 time in total.

Retro-Gamer 😀 ...on different machines

Reply 106 of 119, by vvbee

User metadata
Rank Oldbie
Rank
Oldbie

Claude wrote a microcode fix for the card's hardware T&L glitch that came up for example in 3DMark2000's terrain. Very cool. Fairly involved too, I think near zero chance this would ever have been fixed by hand.

The attachment 3dm2kcmcf.jpg is no longer available

I can still see an occasional issue, but starting to look like proper full hw T&L on this card. Needs to be tested on a slow CPU though to see what the actual value is.

Reply 107 of 119, by vvbee

User metadata
Rank Oldbie
Rank
Oldbie
dr.zeissler wrote on 2026-11-17, 08:16:
If you are comparing G400 to G450 / G550 these things came to my mind. […]
Show full quote

If you are comparing G400 to G450 / G550 these things came to my mind.

- Older Games benefit more from the 128Bit memory interface of the G400 than the DDR-Ram of the later cards (450/550) with 64Bit interface
- G450 does support EMBM in Expendable by RAGE, G550 does not work.
- G450/550 both lack VSYNC in OpenGL mode that is essentiel for lot's of the remakes of classic intro's and cracktros of the 16Bit area (Atari/Amiga)
- G450 has a working Win3x driver for WinOS/2 but no native Win3x driver itself. Don't know if the G550 works too. G400 should have a native Win3x driver.

The second texture-unit of the G550 does not benefit that much in real performance. In short the G550 is mostly only 20-30% faster, but perhaps this depends on the game (engine).

Doc

1: I've seen this claim but not evidence for it.
2: The G550 supports EMBM when using analog output.
3: Both should have vsync control in OpenGL. On: "User3D.SwapIntervalExtFlags"=dword:00000001, "User3D.SwapIntervalExtEnable"=dword:00000000. Off: "User3D.SwapIntervalExtFlags"=dword:00000009, "User3D.SwapIntervalExtEnable"=dword:00000000. Under HKEY_LOCAL_MACHINE\SOFTWARE\Matrox\PowerDesk\Current Settings.
4: No idea.

The G550 benefits considerably when games use multitexturing; it's 40% faster than the G400 in Thief for example. By comparison the G400 MAX is something like 15% faster than the G400 generally speaking, and this is the frame the G550 sits in.

Reply 108 of 119, by dr.zeissler

User metadata
Rank l33t
Rank
l33t

2. Perhaps a software issue in expendable? Have you tried it. It does not work with G550, with G450 everything is fine.
3. Might also be software (engine) related, but it did not work for me.

Retro-Gamer 😀 ...on different machines

Reply 109 of 119, by vvbee

User metadata
Rank Oldbie
Rank
Oldbie

Vsync works for me in the Quakes. EMBM works on my G550. By default the G550 doesn't filter bumpmaps, with driver 6.83 change g550dd32.dll's byte 0x124b0 from 00 to ff. Maybe Matrox had a good reason to do this and maybe not.

The attachment ebrmbe.jpg is no longer available

Reply 110 of 119, by dr.zeissler

User metadata
Rank l33t
Rank
l33t

Please try:
https://cyberpingui.pages-perso.free.fr/realpcremakes.htm
direct link https://cyberpingui.pages-perso.free.fr/dl.php?id=714

Maybe it's just "engine related". I have to check it again. I have to lookup the machine and the driver I was using.
In my case neither G450 AGP or G550AGP had working VSYNC in openGL with the demo scene-remakes that use OpenGL.
But the tearing was not that bad.

Retro-Gamer 😀 ...on different machines

Reply 111 of 119, by vvbee

User metadata
Rank Oldbie
Rank
Oldbie

Run one of the Quake games instead. If it works then you did it correctly and you have the answer. Don't say it doesn't work without testing it though. I used driver 6.83.

The hardware T&L unit on the G550 is apparently a MIPS processor, so you can run arbitrary code on it. I had Claude write a plasma fire effect and patch it in, and yes, it appears to run on the thing. Claude's estimate based on the FPS and the amount of data and whatever was around 120 MHz, so maybe it's running at the speed of the core at 133 MHz or so.

Reply 112 of 119, by Putas

User metadata
Rank Oldbie
Rank
Oldbie

Amazing times we live in. Can Claude or similar tools give more insight into the WARP engine as well?

Reply 113 of 119, by Ozzuneoj

User metadata
Rank l33t
Rank
l33t
vvbee wrote on 2026-11-17, 16:42:

Claude wrote a microcode fix for the card's hardware T&L glitch that came up for example in 3DMark2000's terrain. Very cool. Fairly involved too, I think near zero chance this would ever have been fixed by hand.

The attachment 3dm2kcmcf.jpg is no longer available

I can still see an occasional issue, but starting to look like proper full hw T&L on this card. Needs to be tested on a slow CPU though to see what the actual value is.

Wow, that's very cool. I know this is a whole different issue, but it makes me wonder if there's a way to create a workaround for the broken T&L on the S3 Savage 2000.

Now for some blitting from the back buffer.

Reply 114 of 119, by fosterwj03

User metadata
Rank Oldbie
Rank
Oldbie
dr.zeissler wrote on 2026-11-17, 08:16:
If you are comparing G400 to G450 / G550 these things came to my mind. […]
Show full quote

If you are comparing G400 to G450 / G550 these things came to my mind.

- Older Games benefit more from the 128Bit memory interface of the G400 than the DDR-Ram of the later cards (450/550) with 64Bit interface
- G450 does support EMBM in Expendable by RAGE, G550 does not work.
- G450/550 both lack VSYNC in OpenGL mode that is essentiel for lot's of the remakes of classic intro's and cracktros of the 16Bit area (Atari/Amiga)
- G450 has a working Win3x driver for WinOS/2 but no native Win3x driver itself. Don't know if the G550 works too. G400 should have a native Win3x driver.

The second texture-unit of the G550 does not benefit that much in real performance. In short the G550 is mostly only 20-30% faster, but perhaps this depends on the game (engine).

Doc

The G550 (to include the PCIe version) works in OS/2 with the driver meant for the G450. No idea why Matrox didn't bother doing the same for Windows 3.x and NT 3.51.

Reply 115 of 119, by vvbee

User metadata
Rank Oldbie
Rank
Oldbie

Testing the benefit of hw T&L on the G550 with slower CPUs.

K6-2+/500 (100 MHz), vs the G400:

The attachment a9b190d1e09e.png is no longer available

K6-2/300 (66 MHz):

The attachment b5ec35d9edef.png is no longer available

All in 640 x 480, 16-bit.

The T&L unit on the G550 is fast enough to benefit the K6-2 class CPUs. With software T&L, CPU overhead with the G400 and the G550 is similar except in Quake 3, but with hardware T&L enabled the G550 is much faster.

The G550's hw T&L isn't available so easily in OpenGL, but Claude wrote a MiniGL that wraps into DX7 which then gets the hw T&L benefit. It took about 15-20 minutes to write the MiniGL for GLQuake, then about 10 minutes to adapt it for Quake 3, including time to set up the environment, build the game from source, create a test harness, etc. Pretty crazy for software development. The rendered image in DX7 looks the same as in OpenGL, but I haven't bothered personally comparing side by side, pixel by pixel to see if any rendering work is being skipped.

I don't know if there's much to the WARP engine to know? Claude says the cards have two of them and each processes every other triangle. They get screen space triangles and there's some but limited programmability.

Reply 116 of 119, by mwdmeyer

User metadata
Rank Oldbie
Rank
Oldbie

Wow that is awesome. I remember the Geforce 265 T&L engine wasn't super fast, so faster P3s could match it, the Geforce 2 was a lot better. I wonder how the G550 compares.

Vogons Wiki - http://vogonswiki.com

Reply 117 of 119, by Putas

User metadata
Rank Oldbie
Rank
Oldbie

Amazing, thanks for all the tests. I asked about WARP because it was rumoured to be a RISC-like CPU.
Not the strongest TnL unit, perhaps Matrox was clever to delegate it to fox duties.

Reply 118 of 119, by vvbee

User metadata
Rank Oldbie
Rank
Oldbie

Vs the GeForce2 MX, standing in for the 256, on the K6-2+/500. Using the 30.82 drivers.

The attachment 2797b98bc3a5.png is no longer available

The GeForce2 looks very CPU limited, so the G550 ends up having the stronger hw T&L on this platform. I ran the tests at 640 x 480, 16-bit, but briefly tested higher and the G550 was still faster at 1024 x 768. Maybe the 256 might have less CPU overhead with older drivers.

It's easy to see how the G800 would've been twice the speed of the G400, take the G550 as-is, finish implementing the hw T&L, and keep a 128-bit bus with DDR.

Claude calls WARP "a pair of fixed-function-oriented, microcoded VLIW setup processors," based on what the drivers are doing and some other info.

Reply 119 of 119, by swaaye

User metadata
Rank l33t++
Rank
l33t++

It is amazing to be able to do this kind of analysis now. What about things like what kind of CPU optimizations are present and uncovering game/app profiles. The driver might also have G800 related code in it. Maybe it would be possible to learn more about the progression of G200, G400, G450 and G550.