Yay, more benchmarks! Wheee!
Check out http://www.idsoftware.com/bwh/benchmarks.html for the numbers in tabulated HTML form. Thanks again goes out to Eric Webb, our trusty in-house benchmarker, for doing the tiring task of swapping cards, installing drivers, writing stuff down, rinse, repeat.
The new additions are throughput tests on an AST machine with an Intel Celeron 400MHz CPU. We didn't include the 32-bpp or 1280x1024 numbers since those numbers likely won't change appreciably between a Celeron and PentiumIII (since they're effectively differences in the graphics accelerator, not the CPU).
The truly relevant scores for us are:
- Celeron 400 @ 640x480x16bpp
- PIII/500 @ 640x480x16bpp
- PIII/500 @ 1280x1024x16bpp
- PIII/500 @ 1280x1024x32bpp
We also have some new additions, including a Riva128ZX, Savage4, V3 2K, and a Matrox G400 MAX. We didn't test the ZX or V3 2K on the PIII/500 due to time constraints.
From this data we can gather the following interesting information:
- performance penalty for 16-bit vs. 32-bit for each accelerator
- whether hardware is being saturated, and by inference, whether a driver could use more optimization
If dlights are disabled, then the V3 actually pulls ahead of the Savage4, likely due to faster texture upload performance on the V3 and its drivers.
Performance Hit for 32-Bit
--------------------------
Rage128 @ 640x480: 12%
Rage128 @ 1280x1024: 28%
RivaTNT @ 640x480: 12%
RivaTNT @ 1280x1024: 48%
RivaTNT2 @ 640x480: 4%
RivaTNT2 @ 1280x1024: 40%
Matrox G200 @ 640x480: 0% (driver maxed out?)
Matrox G200 @ 1280x1024: 28%
Matrox G400 MAX @ 1280x1024:13% (16-bit Z-buffer)
S3 Savage3D @ 640x480: 36%
The difference in cost for 32-bit at the different resolutions has to do with the different bottlenecks encountered at each resolution. At 640x480, you're looking at driver/CPU limitations. At 1280x1024, you're pretty much looking at pure graphics hardware limitations. The 1280x1024 numbers give you a good indicator of the absolute cost of 32-bit rendering that the hardware incurs. The 640x480 numbers give you the real world cost if you're interested in better visual quality but don't know how much performance you'll be giving up. At 640x480 the performance loss isn't that huge because much of the time being spent is feeding triangles to the accelerator.
Performance Gain from Celeron 400 to PIII 500
---------------------------------------------
These numbers are fairly easy to interpret. Since, ideally, performance would scale by ~25% moving from a 400MHz CPU to a 500MHz CPU, we'll be able to make some assumptions based on the actual performance difference. If the performance increase is LOWER than 25%, this implies that the hardware is saturated. If the performance boost is GREATER than 25%, this implies that the drivers have some room for optimization on the Celeron and/or that they are tuned for SSE instructions on the PIII.
Note that these are tested at 640x480x16bpp, since this mode emphasizes throughput.
RivaTNT2 +33.3%
G400 MAX +32.7%
RivaTNT +31.6%
Voodoo2 +30.3%
Rage128 +16.4%
Voodoo3 3K +9.8%
S3 Savage4 +8.6%
Voodoo Banshee +6.8%
G200 +6.0%
V2200 +5.3%
RagePro +2.6%
S3 Savage3D +2.3%
Intel i740 +2.0%
So the faster hardware, in general, clusters near the top, since it doesn't get saturated by the slower Celeron 400. So hardware near the bottom of the list is pretty much screaming for mommy even on a Celeron 400, hardware near the top of the list is begging for more.
The Voodoo2 number is anomalous, so I'm going to retest that at some point, and I also hope to put up numbers for our SGI and Intergraph workstations.
General Observations
--------------------
Holy Mother of Moses, did you see those G400 MAX scores at 1280x1024?! That is pure devastation in fill rate the likes we have not seen in years. Now, granted, they are running with a 16-bit Z-buffer, but still...those are pretty awe inspiring numbers.
Now, the flip side is that they're NOT the fastest at 640x480, which means that their throughput (i.e. drivers) could use some work.
The Savage4 made a good showing, however we have the caveat that this was using pre-production drivers delivered directly from S3. We hope that they will release up to date drivers soon for the Savage4. Also, the Savage4 numbers at 32-bit were using a 16-bit Z-buffer, so they had a slight advantage there also.
TNT2 didn't take top honors in many of the fields, just the 640x480x32bpp numbers, however it has the distinction of being very close to the top in nearly all tests, so it's probably the best all around board if you care about high res and 32bpp. TNT still is putting a way respectable showing, and I've seen those Creative TNT boards now for as low as $79. That is as close to a no-brainer as I've seen in a long time.
Voodoo2/3 is still putting in a very admirable show, albeit only in low res modes (Voodoo2) and 16-bit. While its performance is above that of the TNT2 in the 16bpp modes, it's not so far ahead as to have a dominating lead. And it still pales in comparison to the G400 MAX.
We received some TNT2 Ultra boards, but I couldn't get them to work. When those fire up for us I'll post the numbers, and maybe the landscape will change.
Those G400 MAX numbers are high enough that I'm nervous about them, as in maybe I did something wrong, but so far I've run the tests on several machines, and it's pretty difficult to cheat with your drivers when you're fill rate bound.
