TurboGL commonly slower than ICD in Quake 3?

[Ruslan73]

Last days I've tested last drivers 5.52 with and without TurboGL 1.3.

My specs:

hard:PIII-450(L2 ECC off)/128M RAM PC100(CAS3)/G400 DH 32/HDD IBM 10G 7200rpm/Chaintech 6BTM/no overclocking

soft:Win98SE/DX7.0

When tested TGL apperture size was 256M
When tested ICD apperture size was 64M

The results I've obtained:

Q3 Arena Final 1.16n
High Quality (all standard settings except resolution, A3D off, low sound quality)

/s_initsound 0
/timedemo 1
/demo demo001

Results (fps without/fps with TGL)

41.3/35.2 in 800x600x32
28.6/22.2 in 1024x768x32
18.0/12.8 in 1280x1024x32

Looks like standard ICD works faster than TGL in all resolutions in 32-bit color (except 640x480, which I never played). Tomorrow will try 16-bit modes.

This looks very strange 2 me because, as I remember, TGL 1.3 was released AFTER 5.52 drivers.

But may be I'm doing something wrong, or 128M RAM is not enough 4 TurboGL to show all it's strength? Or may be TurboGL intended to be faster than ICD only with G400 MAX?

Any comments?

[This message has been edited by Ruslan73 (edited 22 April 2000).]

[fds]

Something's wrong with those results.

My TGL timedemo scores are either about the same as with the ICD or a few fps higher.

Not a big difference.

There is a rather important difference though. While with the 5.5x ICD I would be in for frequent crashes (especially after several game restarts, and mostly in Half-Life TFC, not Q3), the TGL just works. Well, other than the fact it disables Alt-Tab.

(I would never use 16 bit on a G400.)

[fds]

Just noticed you measured with "all standard settings."

With such settings, even your ICD scores leave much to be desired. I do have a P3-500, but that should not be so much of a difference. I always use 800x600x32, all settings on the highest (some higher than the "high quality" option sets them) except for using Bilinear instead of Trilinear.
Audio qual. is also on high.

On the G400 Max I get about 54 fps and on the G400 OEM SH 32MB I get about 49.

[xortam]

Ruslan73, What's with the disabling of the L2 cache?

[Xavier]

I'm having the same problem with my Turbo GL, with the 5.52 drivers I find the standard OpenGL faster than the Turbo in Half-Life, Tribes, Quake2, usually in 1024x768x16. Any suggestions as to why that might be?

PIII 450 @ 557 Mhz
128 PC 100 SDRAM
7.0G HD
G400 32Mg OEM SH
Sound Blaster PCI 128

[paulcs]

As I understand it, this is not really a problem. TurboGL is optimized to be used at lower resolutions for situations when speed is a premium (for instance, online play).

In addition, the 5.5X ICD is faster than previous ICD's, and has caught up, in many situations, with TurboGL.

Paul
paulcs@flashcom.net

[paulcs]

I take part of it back. The 5.52 ICD, at least on the system I just tested it on, is just a bit faster than TurboGL 1.3 at 800x600, 16-bit, as well.

I get some visual anomalies with the ICD, and I guess that's the trade off at this point.

If you can use the newest drivers, just use the ICD.

Paul
paulcs@flashcom.net

[fds]

At 1024x768x16, the TurboGL being slightly slower than the regular ICD is no surprise.
Ruslan73, however, is saying his TGL is significantly slower even in 800x600x32 (where it should be slightly faster).

If it the ICD works for you stable enough, there is no reason not to use it -- you even get the benefit of Alt-Tab working again.

I'm often playing Half-Life TFC and the ICD just wasn't reliable enough, frequently crashing the whole machine when entering/leaving a game. Unwanted and unplanned reboots annoy the hell out of me. Especially since I have to reestablish my connection to the net.

[Wombat]

Lately some of us have noticed that having a 256MB aperature seems to slow the system down. Try the tests using the same aperature size.

[Ruslan73]

2 xortam:

When writing L2 ECC off I mean that L2 cache is on, but ECC feature of this cache is off. This makes my P3 faster abit, esp. when playing DVD with software decoder.

2ALL:
As about ICD 5.2 vs TurboGL, I have tested 16-bit modes yeasterday and smth. new revealed. In low resolutions (<=800x600x16) TurboGL is much faster than ICD. But in high resolutions the difference is minimal.

After all I've tried 2 overclock my G400 up to 146/195 and tested all modes again. Fps increased, but ICD remains faster in 32 modes on high resolutions.

Q3 Normal (16-bit) P3:450/128/Aperture=256/SwapFile=320M
No overclocking ICD/TGL:
51.4/56.6 640x480x16
49.8/54.2 800x600x16
40.4/40.7 1024x768x16
25.6/25.5 1280x1024x16

Overclocked to 146/195 (chip/memory)
16-bit
38.2/56.6 640x480x16
37.5/55.7 800x600x16
35.7/47.7 1024x768x16
29.2/30.7 1280x1024x16

High Quality (32-bit)
50.4/56.6 640x480x32
44.2/42.1 800x600x32
28.8/27.0 1024x768x32
17.9/16.5 1280x1024x32

Today, I've tested HQ 32-bit o/c 146/195 with biliniar filtering and results in 1024x768x32 showed that ICD's still better:

33.2/29.4 1024x768x32 (HQ+biliniar)

so seems like ICD on my machine still faster than TurboGL in 32-bit modes.

Looks like the main problem with ICD is instability of results : less than 40fps with o/c in 640x480x16 mode and more than 50fps WITHOUT o/c in 640x480x16 mode. (Seems like Matrox programmers still have some work 2 do.)

Also I've experinced some minor visual artefacts with ICD (only). In contrast 2 ICD, TurboGL results were stable and predictable. So, looks like there is some balance here.

As about apperture size - I've not found any difference after I've increased my swap file size to 320M (size of swap is fixed, file is defragmented and moved 2 the beginning of the disk).

Thanks 2 all 4 your response.

[This message has been edited by Ruslan73 (edited 27 April 2000).]

[paulcs]

Hmmm. A fix?

Paul
paulcs@flashcom.net

[Ruslan73]

2fds:
Triliniar filtering slows system about 10% or more in 32bit modes. I remeber, that I had about 52fps in 800x600x32 with standard HQ setting. Today will try with "Highest" quality and biliniar filtering.

But the fact, that ICD is faster than TurboGL in 32-bit modes still makes me sad