Introduction
Banshee, the last chip that 3Dfx released, could not really impress any of us when it was released last year and since Comdex 1998 we are highly anticipating the Voodoo3, a chip that’s supposed to prove 3Dfx’s leadership in the world of 3D-acceleration. The specs sounded good, 183 MHz chip and memory clock for the Voodoo3 3000, two pipelines, 350 MHz RAMDAC, optional digital out and fast 2D. Now we’ve finally reached the time when 3Dfx is giving out boards for us performance-hungry press people and we can see how far this Voodoo3 lives up to our expectations.
This picture shows the final Voodoo3 board, not the sample that I received.
Specs Were Changed Last Week
It’s only a few weeks ago when 3Dfx announced that there will be three different Voodoo3 chips or cards, the Voodoo3 2000 will now be running at 143 MHz and not at 125 MHz as anounced at Comdex 1998, Voodoo3 3000 will now come with reduced specs, 166 MHz for chip and memory clock while Voodoo3 3500 takes over the specs of the previously announced Voodoo3 3000, with 183 MHz chip and memory clock. This seems a bit odd already, making it look as if 3Dfx is not getting a good yield of 183 MHz parts, so that the majority of Voodoo3 chips will only run at the 166 MHz clock and the 183 MHz chips will only be available in limited numbers. 3Dfx is also facing the fact that there’s hardly any memory that runs at 183 MHz, which complicates things additionally.
STB Now the Only Card Maker of Voodoo3-Cards in the US
A lot has changed within 3Dfx in the last months, they acquired STB and are now selling their own cards with STB’s card manufacturing plants. This upset many of the former ‘lovers’ of 3Dfx, particularly Creative and Diamond, who will now put a lot of effort into pushing NVIDIA, since this is the only chance they’ve really got on the 3D-card market. The story looks a bit different in Europe, where STB is hardly known and where 3Dfx depends on other card makers to market their new chips. However, why should any European 3D-card maker be crazy enough to help 3Dfx now, when 3Dfx is anyway planning to take care of the difficult, but big European market by themselves as soon as they can?
The situation is clear, 3Dfx needs the Voodoo3 to be a success, otherwise they’ll be in serious trouble. Voodoo3 needs to be a real performance leader, since Banshee was never up to impress any of us performance hungry hardware freaks or hardcore gamers, although it may have sold reasonably well. 3Dfx is at risk of losing their reputation of providing the best and fastest 3D chips and Voodoo3 really needs to be what Banshee never has been. This preview will try to answer the question if the performance of the new Voodoo3 will be good enough for 3Dfx to stay on top of things.
The Voodoo3 Board
The board I received is a Voodoo3 3500. The PCB is black and thus looking as if it was a Quantum3D-part. Quantum3D has the tradition of only making cards with black PCBs. The chip as well as the memory clock was indeed 183 MHz, the chip is covered by a pretty large heat sink and fan combo and the memory used on the card is 16 MB of 6ns SGRAM from EtronTech. I cannot complain about any excessive chip temperature when running 3D, the fan seems to take good care of Voodoo3’s heat production. Unfortunately I couldn’t use my thermal probe, because I haven’t got one in the US-lab yet. The board runs absolutely stable at the 183 MHz. It features a S-VHS video out that’s working fine and a 20-pin digital out driven by the onboard ‘LCDfx’-chip. Unfortunately I couldn’t test the digital out yet.
The Driver
The driver provided was from February 10, 1999, driver version 4.11.01.0409-1.00. This includes Glide2-driver version 2.60.00.0382, Glide3-driver 3.10.00.0381and Mini-VDD version 4.11.01.0409-1.00. I was surprised enough to see that the driver file calls itself ‘3Dfx Banshee display driver’ when you check the version and after comparing this V3-driver with the latest Banshee driver, I found out that the files are looking the same, although they aren’t identical. It seems a matter of fact that Voodoo3 uses drivers that are only slightly modified Banshee-drivers, which raises the question if Voodoo3 is nothing but a faster clocked and double pipelined Banshee after all. Unfortunately I didn’t try to run Banshee with those drivers or Voodoo3 with the original Banshee-drivers, but I’ll do that as soon as I’m in my lab again.
3Dfx has still not been able to provide a real OpenGL-ICD with the Voodoo3-drivers, so that you have to use the Mini-GL of Banshee if you want to play Quake2 or any other OpenGL-based game. Without an ICD there will be no support for professional OpenGL-applications. It seems as if 3Dfx takes even longer than Matrox to finish their long overdue OpenGL-ICD, which I consider as pretty embarrassing for both companies. NVIDIA, ATI, 3DLabs and S3 are much more advanced here, they all finished their ICD quite a while ago.
Problems
I had a really hard time running Quake2 at resolutions of above 1024×768, because often only one third of the screen was visible, the rest remained black. There was a particular problem at 1152×864, where the picture produced by the Voodoo3 was never able to fill the whole monitor screen. I managed to run Quake2 at 1600×1200 and 1280×960 only after switching the desktop resolution to 1600×1200 in the first place before starting Quake2. In Shogo the Voodoo3 let’s you choose a resolution of 1600×1200 also, but none of the moving objects in the game will be textured anymore if you run it at this resolution, which looks pretty horrible of course. I guess the Voodoo3’s lack of AGP-texture support is mainly responsible for that behavior. The only other problem that I found was Turok2, which would crash immediately after I started it, way before I could choose between Glide or D3D. I found out that this was due to the Glide3-driver coming with the drivers I had, which was of revision 3.10.00.0381. After replacing it with glide3x.dll ver. 3.10.00.0363 out of the driver set from January 22, 1999 Turok2 ran fine.
Image Quality
My super 21″-monitor from Sony did unfortunately not travel with me to the US, so that I could not yet test the Voodoo3 with this reference CRT. I will do that as soon as I’ve returned to Europe. The 3D-image quality was fine in all the games I ran. As a matter of fact, I couldn’t find any obvious difference between Voodoo3, NVIDIA RIVA TNT and ATI Rage128 at 16 bit rendering. We all know that Voodoo3 cannot do real 32bit rendering, so that I didn’t do any comparison here.
Test Systems
Pentium III and Celeron System
- Asus P2B motherboard with Intel 440BX-chipset
- 128 MB PC100 SGRAM
- Adaptec 2940U2W SCSI Host Adapter
- IBM DGVS-09U ultra wide SCSI hard drive
K6-2 and K6-3 System
- Asus P5A motherboard with Ali Aladdin V chipset
- 128 MB PC100 SGRAM
- Adaptec 2940U2W SCSI Host Adapter
- IBM DGVS-09U ultra wide SCSI hard drive
Winstone99 was ran at 1024x768x16bit and 1024x768x32bit, 85 Hz refresh rate.
‘VSYNC’ was turned ‘off’ in all of the 3D-benchmarks.
Driver Revisions Used
3Dfx Voodoo3 |
4.11.01.0409-1.00 |
3Dfx Voodoo2 |
4.11.01.0350-2.18 |
ATI Rage128 |
4.11.6061 |
NVIDIA RIVA TNT |
4.10.01.0106 |
2D Performance
I ran Winstone99 with Voodoo3, RIVA TNT and Rage 128 at 16 bit as well as at 32bit color. The results weren’t even worth making a chart. If you run the tests about three times on each graphics card you will find out that at 16 bit as well as at 32 bit desktop color depth the Winstone99 results are absolutely identical. Voodoo3 is as fast as the other two top 2D-performers, but it’s the only one with a RAMDAC of 350 MHz, which enables highest desktop resolutions on high-end monitors. For me with my excellent Sony-monitor this is an important aspect, which is why I won’t hesitate putting a Voodoo3 into my system.
3D Performance
Before I started comparing Voodoo3 with other graphics card, I decided to compare the different versions of Voodoo3 between one another. To get a good measure of the performance of Voodoo3 2000 and Voodoo3 3000 I ‘underclocked’ my Voodoo3 3500 to lower chip clocks. However, I did not ‘underclock’ the memory clock as well, so that the results of the 2000 and the 3000-version displayed here will be a little bit higher than what a real Voodoo3 2000 or 3000 will score. The difference is not too significant however.
The Three Different Voodoo3s – Quake2 Crusher
The first benchmark I used was Quake2 and the Crusher.dm2 demo from Brett ‘Three Fingers’ Jacobs. This demo is pretty tough and thus not showing as high frame rates as ‘demo1’. However, it is a much better measure of the real performance a system offers, because it’s an example for a worst case scenario. The frame rate in Quake2 should hardly ever drop underneath the result scored in ‘Crusher’. This makes ‘Crusher’ the best benchmark for the hard core gamer, who doesn’t want to know how high the frame rate can go, but how low it can drop.
Quake2 Crusher Frame Rate [fps] |
3Dfx Voodoo3 3500 |
3Dfx Voodoo3 3000 |
3Dfx Voodoo3 2000 |
1600×1200 |
25.1 |
22.6 |
19.5 |
1280×960 |
35.2 |
33 |
28.6 |
1152×864 |
40.6 |
37.4 |
32.9 |
1024×768 |
46.4 |
43.4 |
39.0 |
960×720 |
48.2 |
46.5 |
41.7 |
800×600 |
51.9 |
51.9 |
49.4 |
640×480 |
52.5 |
52.1 |
52.0 |
Crusher is so tough that the frame rate never excels 53 fps. Voodoo3 3000 is not a whole lot slower than Voodoo3 3500 and this is measured in a Pentium III 500-system. The CPU-limitation is of course even higher with a slower CPU, so that this test makes you wonder if the Voodoo3 3500 is that much of an advancement over Voodoo3 3000 after all. Voodoo3 2000 performs significantly worse than its two bigger brothers.
The Three Different Voodoo3s – Shogo Tomsdemo
The next benchmark I used is less well known in the gaming community. On the look for a high quality Direct3D game that is using as many new DirectX6-features as possible I came across Shogo again. This time I sat down for a whole day to figure out how to make a decent demo file. I finally succeeded and recorded ‘tomsdemo’, which you can download from my ftp-site. This demo runs a lot better than the ‘reverend’-demo from ‘3 Fingers’ site, which is pretty useless for benchmarking. The new ‘tomsdemo’ could become the reference test for D3D-performance in the next few months.
Shogo Tomsdemo Frame Rate [fps] |
3Dfx Voodoo3 3500 |
3Dfx Voodoo3 3000 |
3Dfx Voodoo3 2000 |
1280×1024 |
30.4 |
27.5 |
24.5 |
1152×864 |
39.4 |
35.5 |
30.6 |
1024×768 |
49.3 |
43.4 |
37.3 |
800×600 |
71 |
65.9 |
57.9 |
640×480 |
82.4 |
81 |
75.8 |
The picture is similar to the Crusher-demo, but in Shogo the framerate reaches higher numbers at low resolutions, most likely because ‘tomsdemo’ isn’t quite as tough as Brett Jacob’s excellent Crusher-demo. You can again see that the performance difference between the different Voodoo3 versions correlates perfectly with the chip clock rates of those cards.
Comparison of Voodoo3 with Actual 3D-Cards – Quake2 Crusher
The specs of Voodoo3 can already tell what results we should expect. Rage128 and RIVA TNT will not have a chance against any of the Voodoo3 boards, because none of the two has got the fill rate to stand up against them. Voodoo2 SLI however has got a fill rate of 366 MTexels/s, although only 180 MPixel/s. Games that use multi-texturing should run just as fine on Voodoo2 SLI as on Voodoo3 3500. Let’s see what the results show.
Quake2 Crusher Frame Rate [fps] |
Intel Pentium III 500 |
Intel Celeron 400 |
AMD K6-3 450 |
AMD K6-2 400 |
3Dfx Voodoo3 3500 |
46.4 |
38.9 |
40.2 |
35.7 |
3Dfx Voodoo3 3000 |
43.4 |
37.6 |
39.4 |
35.1 |
3Dfx Voodoo3 2000 |
39.0 |
34.3 |
36.0 |
32.6 |
3Dfx Voodoo2 SLI |
43.5 |
38.2 |
42.3 |
38.2 |
NVIDIA RIVA TNT |
30.9 |
28.7 |
25.3 |
20 |
ATI Rage 128 |
32.6 |
26.2 |
22.7 |
18.5 |
This test shows some really nice results, doesn’t it? First of all, Voodoo2 SLI is indeed pretty damn close to Voodoo3, but not quite as close to Voodoo3 3500. The systems powered by an AMD-CPU however score better with Voodoo2 SLI than with Voodoo3, which I consider as very interesting. You can also see very nicely that all the 3Dfx cards have a really good 3DNow!-support, which is why you find the graph going up again at the stage of the K6-3. ATI and NVIDIA suck at 3DNow!-support, which you can see by the steadily falling graph. It should be said crystal clear, owners of AMD-CPUs should stay away from any graphics card that is not from 3Dfx until the other 3D-chip makers learned their lesson properly.
Comparison of Voodoo3 with Actual 3D-Cards – Shogo Tomsdemo
Shogo has not got an inside 3DNow!-support built in yet. I have tested a beta version of the 3DNow!-renderer, but is was too unstable for testing. The missing 3DNow!-support can be seen when you compare the Shogo with the Quake2 graph. Quake2 does not have an internal 3DNow!-support either actually, as a matter of fact Id couldn’t care less about AMD-support, as proven by a lovely and very polite email that I once received from Grand Master (or is it Emperor of China, or already God himself?) Mr.(or is it ‘Her Majesty’ ?) Brian Hook, stating that Q2 doesn’t have 3DNow!-support. Quake’s 3DNow!-support comes from the 3Dfx-mini-GL and the default OpenGL files, which are developed by AMD itself.
Shogo Tomsdemo Frame Rate [fps] |
Intel Pentium III 500 |
Intel Celeron 400 |
AMD K6-3 450 |
AMD K6-2 400 |
3Dfx Voodoo3 3500 |
49.3 |
47 |
44.6 |
39.8 |
3Dfx Voodoo3 3000 |
44.2 |
43.1 |
42.5 |
38.4 |
3Dfx Voodoo3 2000 |
37.6 |
36.8 |
37.1 |
34.1 |
3Dfx Voodoo2 SLI |
42.8 |
40.9 |
39.7 |
36.9 |
NVIDIA RIVA TNT |
29.8 |
26 |
24.1 |
22.4 |
ATI Rage 128 |
28.4 |
25.8 |
25.3 |
23.8 |
Voodoo2-SLI does not look quite as good under Direct3D as under Glide. If we use Shogo as a good measure for Voodoo3’s Direct3D-performance, then it may be well worth getting a Voodoo3 for Direct3D-games, even if you are owning two Voodoo2-boards already. You can also see that even Voodoo3 2000 is still smoking the competition from ATI and NVIDIA.
AGP-Performance Test with S3’s Quake2 Mon2-Demo
The Mon2-demo I once got from S3 has proven really very helpful since. S3 supplied me with this demo that uses about 9 MB of textures to show me the texture compression feature of Savage3D, but you can also test if a graphics card can move these textures in and out AGP-memory quickly. So far 3Dfx-products looked really bad in this benchmark, let’s see if Voodoo3 makes any kind of exception.
Many of you may be surprised to see that Voodoo3 is scoring even worse than Voodoo2 SLI in this large textures test. The explanation for this behavior is pretty simple though, Voodoo2 SLI has with 24 MB vs. 16 MB exactly 50% more onboard memory than Voodoo3 to store some of the textures. This gets even worse when Voodoo3 is using triple buffering, something that Voodoo2 isn’t even able to do at all. It’s also interesting to see that ATI seems to have a problem with their OpenGL-ICD, which doesn’t seem to support the AGP-texturing that’s certainly supported by the Rage128-chip. TNT is easily smoking the whole competition, its AGP-support works under OpenGL as well as under Direct3D. I wonder if it is all right with you, that even 3Dfx’s latest product still ignores AGP-texturing blatantly.
Conclusion
Voodoo3 Advantages |
Voodoo3 Disadvantages |
|
|
I guess you want to know now if I would recommend Voodoo3. As a matter of fact I do. However, I don’t recommend the purchase of Voodoo3 because it’s really satisfying my expectations, I recommend it because there’s currently nothing better out there. I would also like to criticize 3Dfx’s ambitious claims Voodoo3 would run games at no less than 60 fps at 1280×1024. Have a look at my benchmarks and check if Voodoo3 was anywhere near 60 fps at resolution of even less than 1024×768. Voodoo3’s life spawn won’t be as long as the life spawn of Voodoo1 or Voodoo2. It won’t be long until a competitor has a product ready with the same performance but with support of AGP-texturing and support of 32-bit rendering. So I would like to commend 3Dfx for a job well done for the time being, Voodoo3 is faster than any other 2D/3D-solution and it can surpass Voodoo2 in most situations as well. Even owners of two Voodoo2-cards should consider Voodoo3, because the Voodoo3 offers a definitely superior 3D-image quality and higher resolutions that are certainly playable, although not at 60 fps.
All in all Voodoo3 is a decent product and can offer a lot of fun to most of the 3D-gamers. It’s not a big step forward though, which I consider as a bit disappointing. 3Dfx got away with it for now, but times will get a lot harder for them pretty soon.