Introduction
3D-gaming used to be a tough call with non-Intel CPUs for a long time. The floating-point performance of CPUs from AMD, Cyrix, IBM, IDT, etc. could never really come close to the number crunching ability of a Celeron or Pentium II. Last year AMD finally did something against this and introduced their own ‘streaming SIMD’-solution called ‘3DNow!’. This extension to the CPU instruction set was particularly targeted to make AMD-processors perform better in 3D-games and other applications that require floating point calculations on arrays of numbers. It has been taking a while until software developers optimized their games or other applications for 3DNow!, but today there’s quite a variety of titles you can get with 3DNow!-optimization. 3Dfx used to be the first hardware developer who optimized their drivers for 3DNow! as well. This is why so far every owner of a K6-2 or K6-3-system was trying to get a 3Dfx-product, if he wanted to do 3D-gaming. Now NVIDIA claims to have optimized their TNT2-drivers for 3Dnow! as well. The question underlying this article is pretty simple: Does any new 3D-chip perform better than Voodoo3 under 3DNow! ? We found the answer to that question for you, testing products from 3Dfx, NVIDIA, Matrox, S3 and ATi.
Platform Compatibility
Unfortunately, we’ve been working on this article for several weeks now. What should have taken several days turned out to be quite time consuming. Getting an Nvidia TNT2 based board to work in an ALI-V based motherboard along with getting a Voodoo3 clocked at 183MHz to work on the FIC PA-2013 with an AMD-K6 III 450MHz was nothing more than amusing.
Issues with Nvidia TNT & TNT2 and Aladdin V based motherboards
Most everyone has heard the nightmare regarding TNT & TNT2 working with the ALI-V chipset. Up to date it has always been a hit or miss regarding the compatibility of this combination. Typically, with this configuration the unlucky end-user experiences either intermittent lockups when running 3D games or the familiar “won’t boot Windows” problem. This problem lies within the ALI Aladdin V chipset and some timing issues with the ALI-V AGP bus. Unfortunate for Nvidia, this timing issue seems only to arise with TNT & TNT2 based video boards. Both Nvidia and ALI have been working together to resolve these problems. This problem seems to be corrected with a new AGP driver from ALI and a new video driver from Nvidia on some ALI-V based platforms. We have an ASUS P5A revision 1.04 with works great with ALI’s v1.60 AGP driver along with Nvidia’s version 1.88 driver. However, our Microstar MS-5169 version 2.1 (ALI-V based) with the same drivers doesn’t work. Hopefully, both Nvidia and ALI will be able to provide a single set of drivers that takes care of these issues on all ALI based platform. We’ll be sure to pass this information along once we receive it.
Issues with 3Dfx Voodoo3 and FIC PA-2013 motherboard
Because of the known issues between TNT & TNT2 in ALI-V based motherboards we thought it would make sense to choose a platform that was happy with the variety of video boards we were going to throw at it. We chose FIC’s PA-2013 with a whopping 2MB of cache on it. Everything seemed to work great until we began testing the Voodoo3 at 166MHz and 183MHz with the AMD-K6 III 450MHz. After several hours of testing both PA-2013 boards began to fail. First there were problems with some intermittent hangs, followed with intermittent reboots, and finally the system wasn’t even stable in the CMOS setup anymore. We contacted FIC and they quickly replaced the bad motherboards but the new ones soon failed after a few 3D tests as well. The FIC PA-2013 board was happily running the AMD-K6 III 450MHz with the ATI Rage 128, Nvidia TNT2, Matrox G400, and S3 Savage4. Also, the PA-2013 and Voodoo3 (143MHz, 166MHz & 183MHz) ran fine as long as the AMD-K6-2 was installed in the board. Six PA-2013 boards later we decided to retest everything in the ASUS P5A revision 1.04 and the new AGP driver from ALI-V (v1.60) as well as the video drivers from Nvidia v1.88. FIC is aware of this problem and promises a solution. They seemed to have narrowed the problem down to a component on the PA-2013’s power supply. In any case, the ASUS P5A (w/512KB cache) out performed the FIC PA-2013 (w/2MB cache) with every video adapter I tested in this review.
Tested Environment
CPU
- AMD-K6-2 400MHz
- AMD-K6 III 450MHz
Motherboard
- ASUS P5A Revision 1.04 (512KB cache)
- BIOS Revision 1007
- AGP Driver Version 1.60
Memory
- 128MB PC-100 SDRAM
Hard Disk
- Western Digital AC418000-00DW (18 MAR 99)
Audio
- Diamond Multimedia Monster Sound MX300
- Driver Version 1.20b
Operating System
- Windows 98
- DirectX 6.1eng
Video Adapter(s)
- 3Dfx Voodoo3 = Driver date 4-27-99 (Quake III Compatible driver), MiniGL version 1.48
- ATI Rage 128 = Driver date 5-25-99
- Matrox G400 MAX = Driver date 5-21-99
- Nvidia TNT2 = Diamond Multimedia Viper V770 Ultra, Driver date 6-4-99
- S3 Savage 4 = Diamond Multimedia Stealth III S540, Driver date 5-26-99
Games
Expendable Demo
- triple buffering enabled
- highest quality for all video settings
- audio disabled (however a sound card must be present to run)
- movies disabled
- vsync sync to refresh disabled
Quake II 3DNow! version 3.20
- CD Audio disabled (+set cd_nocd 1)
- Audio disabled (+set s_initsound 0)
- Joystick disabled (+set in_initjoy 0)
- vsync sync to refresh disabled
You may be surprised that you won’t find different clock speeds of the graphic chips in the results. The reason is quite simple. Expendable as well as the Quake2 Crusher-demo are heavily depending upon the CPU. In case of the K6-2 and K6-3 there’s no difference in the frame rate results of those benchmarks between e.g. a Voodoo3 at 143 MHz or at 200 MHz. You would find differences in some today meaningless benchmarks as Forsaken, Incoming or Quake2-demo1. The tests we did were targeted to the high-end gamer, who is using top games and who considers playing online. In both cases the CPU is working very hard, which is why we chose benchmarks that are tough on the CPU, thus supplying you with realistic and useful numbers.
Direct X Performance using Expendable
I decided to use the Expendable Demo by Rage Software. This game uses Direct3D and has a built-in benchmark that provides data regarding the average frame rate obtained during the playback of the demo. The Expendable game will be used a lot here at Tom’s Hardware Guide since it supports environmental bump mapping, multi-texturing, and both 16bit & 32bit modes. Besides the built-in performance metric the game also has some incredible graphics and is also optimized for 3DNow!. All testing was done at 1024x768x16bpp.
Both Chart I & II show that TNT2 was the fastest while the Matrox G400 barely edges out the Voodoo3 for second position. Still the differences between the first three are very little, it shows that all of them have a good 3DNow!-optimization. It is easy to see and realize that no card is edging out its fill rate or triangle rate limitation in this test. The numbers are purely depending on how much work the drivers need to get done by the CPU for supplying the geometry data. In Direct3D the drivers of TNT2, G400 as well as Voodoo3 are doing a good job on using the special features of 3DNow!.
OpenGL Performance using 3DNow! Quake II version 3.20
Even though the TNT2 was the clear winner in the Direct3D-testing using the Expendable Demo, the same doesn’t hold true for 3DNow! Quake II v3.20. For the Quake II testing we used the ‘industry standard’ CRUSHER.DM2 demo created by Brett Jacobs a.k.a. “3Fingers”. For the Voodoo3 based board we installed the MiniGL v1.48 drivers and renamed the 3DFXGL.DLL file in the QUAKE2 directory to OPENGL32.DLL to take advantage of the 3DNow!-optimization for Voodoo3. All testing was done at 1024x769x16bpp.
Voodoo3 pulls into the lead under 3DNow! Quake II overtaking the TNT2. This is mainly due to the special Glide-driver of Quake2, which is optimized better than the default OpenGL-driver. It’s still interesting to note that the TNT2 sticks up pretty well against Voodoo3 running Glide. Savage4 is surprisingly already edged out, the frame rate scored with K6-2 is identical to the K6-3 frame-rate results. Matrox needs to improve its OpenGL-ICD for G400. The Direct3D-performance of G400 is certainly excellent, however under OpenGL-games it takes a big hit when the CPU is working hard. Rage128 is sticking up quite well actually, particularly if you compare it to the Matrox G400.
Summary
One thing is certainly still true, an owner of a K6-2 or K6-3-system can hardly go wrong with a product from 3Dfx, at least Voodoo2 and Voodoo3 are good performers with an excellent 3DNow!-support. NVIDIA has indeed managed to do an excellent work on their new drivers as well. TNT2 is the best 3DNow!-performer in Direct3D, which is still the platform for the majority of 3D-games. It’s very impressive to see the difference between the old and the new drivers for TNT2. The OpenGL-performance of TNT2 is not quite as high as Voodoo3’s Glide-performance under Quake2, however, it’s not far behind. This story looks very different with Quake3. Id’s upcoming game requires a real OpenGL-ICD, which is the strength of NVIDIA and the weakness of 3Dfx. Once the new Q3test-version is out you will see that TNT2 performs better than Voodoo3 in this OpenGL-game as well. In regards to compatibility, we recommend hanging tight before installing a NVIDIA TNT2 based graphics adapter into an ALI-V based motherboard until both NVIDIA and ALi get all of their drivers working together. If you own an FIC PA-2013 and an AMD-K6 III 450MHz we advise to wait until FIC understands how to fix their onboard power supply before installing a Voodoo3 3500 (183MHz). The real surprise in this comparison was the Matrox G400 MAX. It is not quite true that G400 is simply requiring a lot of CPU-power, it performed very well under Direct3D even with a K6-2 400. Once Matrox can bring the G400-OpenGL-ICD up to speed, we’ll have another tough competitor in the 3D-graphics market for Socket7. 3Dfx do certainly not own this market alone anymore.
In the works at Tom’s Hardware Guide is a complete review of the Matrox G400 using a variety of processors and game titles followed by a complete performance comparison of all the 3D graphics adapters.