G400 SD 128Bit / G450 DDR 64Bit / G550 DDR 64Bit. G4x/5x have working vSync in D3D but not in OpenGL.
G400/450 works in EMBM G550 does not work in Expandable, don't know if it works in others.
G550 has an additional Texture-Unit, but the performance is only slightly better.

It would be interesting to add the G450 to this comparison.

G550 PCIe is that slow even while running native? I was thinking about grabbing one for fun, but G200 level of performance is underwhelming.

Running native? Ithought it has a bridgechip or am i wrong here?

It looks to be slow due to low vertex throughput, slightly higher fillrate than the G400. Passable for older games and may be the most compatible PCIe card you can find for these games. Probably also a strong GUI accelerator like the original G550, has snappy VM performance in any case.

The VFIO enabled fork of 86Box, https://github.com/richardg867/86Box, looks to not have worked with the G550 PCIe at all until about three months ago with commit 6c72fc2. It's a nice one with 1600 files changed so not possible to figure out which change fixed it. But looking at the code I can see there's some VFIO workarounds for certain ATI, Nvidia and Trident cards, so the gist may be that the Matrox needs a workaround as well for artifact free rendering. But implementing it would require specialist knowledge, or just luck.

Could be useful to have passthrough for the G550 in 86Box. It works with QEMU/KVM but any speed sensitive game is going to give trouble with the overly fast CPU.

Tested to see if the G550 PCIe can work in a Windows 3 VM. Someone said the G450 OS/2 driver works with the G550, but that's not the case, so can't use Windows 3 through Win-OS/2. I combined the Windows 3 driver from the G450 OS/2 driver set with the G400 driver set for native Windows 3, but again it's not compatible with the G550. Maybe for the G450 it would work, but that's not a PCIe card.

PCIe vs. AGP version of the G550 in vertex throughput. Showing FPS ratio of the PCIe to the AGP on Y, triangle count on X. Rendering a static textured mesh using Direct3D 7 in Windows 98.

So the PCIe version starts off at 80% of the speed of the AGP version, makes sense since it's clocked at 80%. But with anything more than 500 triangles it takes a massive extra hit. I can't imagine the PCI bus being this gimped vs. AGP so it's more likely the bridge chip, the TI XIO2000. But then I don't know.

In this particular test the PCIe version tops out at about 300000 triangles per second, the AGP version does just over a million.

So the PCIe version should do relatively well in low poly games, and suffer more in games with poor visibility culling. But from what I've seen the polycount probably isn't the only factor.

Could be that I made a mistake on that win3 compatibility, I have to recheck it.
G400 offers native Win3x drivers. G450 does not work with these, but G450 OS2 drivers work with winos2, So you can get high-true-color and hires support on win3(within OS2 of course) with g450. The G550 works with the g450 driver in OS2, but if it really supports Winos2 as well...I am not sure about it anymore. It has to be retested. I removed my G550 card, because it get's too hot in my E600. In the E600 advertisement from FSC there is only the G450 DVI listed as suitable option, not the G550. Also the G450 offers EMBM (expendable), the G550 refuses this. There must be more differences than just an additional TMU on that G550. compared to the G450.

Warp 4 wasn't compatible with the G550, not with the PCIe version passed through to a VM.

Approximating from the earlier data the PCIe version could hang within about 20% of the AGP version up to about 200000 triangles per second, presumably the bridge chip or whatever starts saturating beyond that. At that rate 3000 triangles per frame gets you about 60 FPS, so games like Thief 1 or Quake 2 might run comfortably and see these two cards perform relatively close. That's ignoring various other factors, but a first guess.

So in other words G550 PCIe is just a gimped even further PCI G550. I wonder if AGP G550 without GART driver (i.e. PCI 66 mode) behaves in similar manner or it's unique interaction with PCI-PCIe bridge.

You brought the pessimism but not the data. The PCI version uses a bridge chip also but an older design, so I'll give the PCIe version the benefit of the doubt. It's the more useful version of the two in any case.

I was intending to test earlier drivers with the passthrough PCIe G550 in Windows 98 but looks like they didn't support it. 6.50 boots into a blank screen and 6.71 (first official driver for the G550) doesn't have 3D acceleration at least by default. There are potential performance benefits for certain games in older drivers for the AGP card, but at least the newest drivers works.

GLQuake:

Data for the G450 and Voodoo3 from vgamuseum.info.

The G550 PCIe takes the smallest hit from doubling the resolution, so it's the least fill rate limited. At 1024 x 768 it's on par with the G400, Voodoo3 etc. Keeping in mind it's running in a VM so let's say 10% slower than direct native, and has about 20% lower clocks compared to the G550 AGP.

A few other games:

The PCIe's FPS is as my napkin calculations predicted but even in these lower poly games the AGP version is far ahead. Again the PCIe version isn't primarily limited by fill rate. If it's the bridge chip then patterns of data access could play into it and not just the amount of data.

vvbee wrote on 2025-09-03, 23:45:

GLQuake: […]
Show full quote

GLQuake:

The attachment ec825de9932d.png is no longer available

Data for the G450 and Voodoo3 from vgamuseum.info.

The G550 PCIe takes the smallest hit from doubling the resolution, so it's the least fill rate limited. At 1024 x 768 it's on par with the G400, Voodoo3 etc. Keeping in mind it's running in a VM so let's say 10% slower than direct native, and has about 20% lower clocks compared to the G550 AGP.

A few other games:

The attachment 0e6b00dcee7b.png is no longer available

The PCIe's FPS is as my napkin calculations predicted but even in these lower poly games the AGP version is far ahead. Again the PCIe version isn't primarily limited by fill rate. If it's the bridge chip then patterns of data access could play into it and not just the amount of data.

My experience with G450 is that it has quite good GLQuake performance but it has driver bugs that absolutely destroy its performance in Quake2 unless you disable dynamic lighting.