Introduction
It’s not too long ago that Tom’s Hardware Guide introduced 3Dfx’s upcoming Voodoo3 chip to you, and many of you may have thought that this new chip will ensure 3Dfx’s role as supplier of the world’s fastest 3D accelerators for the future. Nvidia’s upcoming RIVA TNT2 looked like a merely overhauled version of the well known RIVA TNT and although it was clear that TNT2 would be faster than TNT, still only few expected that TNT2 would be a particularly interesting product. I heard about TNT2-Ultra some six weeks ago and for the first time I saw NVIDIA’s chance to kick 3Dfx from its 3D-performance throne. TNT2 is not just a TNT chip shrunk to 0.25 micron and thus simply clocked higher, TNT2 has overall been improved and the clock speeds it’s able to reach are hardly anything short of Voodoo3’s clock speeds. NVIDIA explained to me back in January that TNT2 wouldn’t even need to be clocked as high as Voodoo3 to beat the performance of 3Dfx’s latest chip, they promised me that TNT2 would be faster than Voodoo3 clock for clock. It turns out that NVIDIA kept another promise. The days when 3Dfx was the performance leader in the 3D gaming scene are over now, after no less than 3 very successful years in which 3Dfx earned the highest respect from everyone in the PC area.
The New Features of TNT2
So what has Nvidia done? TNT2 is indeed the brother of TNT, shrunk down to .25 micron, but several other enhancements put TNT2 significantly apart from TNT. First of all, the rendering pipelines were improved, making the TNT2 some 10 – 17% faster than TNT at the same clock speed as shown in the graph below. The next difference is of course TNT2’s ability to run at much higher clock speeds than TNT could, 175 MHz is so far the highest that I have seen, which is almost double the 90 MHz default clock speed of TNT. TNT2 supports AGP 4x, but that’s not too important right now since Intel’s Camino chipset with AGP 4x-support won’t be out until mid 1999. TNT2 comes with a 300 MHz RAMDAC, offering resolutions of up to 2048×1536 at 60 Hz, it can display 1920×1200 at 32 bit color with a 85 Hz refresh rate. The frame buffer size of TNT2 will be 32 MB, double the amount of TNT or Voodoo3 and TNT2 will also have a digital flat panel interface. Those features make TNT2 a much more interesting product than I would have thought a few months ago and a much scarier product for 3Dfx than they would have anticipated.
For this comparison I ran TNT2 as well as TNT with the 0170 driver and ‘underclocked’ TNT2 down to the same clock frequencies as TNT, which are 90 MHz chip clock and 110 MHz memory clock. The improvement over TNT is 17% in Quake2 Crusher and 10% in Shogo Tomsdemo.
The Boards
I was testing 3 different versions of TNT2-boards. A 16 MB board with a TNT2 running at 125 MHz chip clock and 150 MHz memory clock, a 32 MB board with TNT2 running at 150 MHz chip clock and 183 MHz memory clock and finally a special Diamond 32 MB board with TNT2 running at 175 MHz chip clock and no less than 200 MHz memory clock. The Diamond board was using 5 ns SDRAM, the other two boards were using 5.5 ns SDRAM. Those three different clock speeds are simply showing some configurations that are possible with TNT2. None of the 3D-card makers could yet confirm which kind of configuration they are planning to use. I am also deliberately avoiding the term ‘ultra’ for the higher clock speed versions of TNT2, because NVIDIA is not sure if the term ‘ultra’ will be used to distinguish between the different speed versions. NVIDIA expects to produce a pretty high yield of parts running at 150 MHz or above, they are suggesting numbers in the two-digit percentage range. NVIDIA leaves it up to the card vendor to decide on how high to clock TNT2 on their boards. The two reference boards were equipped with a S-VHS video-out, the Diamond board came without it. None of the boards had a digital output for flat panels. All boards came with a cooling fan on top of the TNT2-chip, which did a very good job on keeping TNT2 pretty cool. The TNT-board from STB is e.g. getting seriously hotter than any of the TNT2-boards did.
The Drivers
TNT2’s new driver is an improved version of NVIDIA’s latest ‘Detonator’ driver for TNT. TNT2 runs with TNT-drivers as well, but it shows better performance with its own drivers. TNT does run with the TNT2 driver as well, but it’s very unstable. I guess that the TNT2-driver takes advantage of TNT2’s improved rendering pipeline timing, which causes problems when using it with a TNT-board. The driver used for the testing was rev. 0170. I just received the new driver 0172b now, which is overall a bit faster and has improved 3DNow!-optimizations, but it came to late to re-run all the benchmarks before publishing. Keep this in mind when reading the results, the new driver makes TNT2 even a bit faster, particularly with K6-2 and K6-3. We should also not forget that this here is a preview, we can expect improved performance once TNT2 starts shipping. NVIDIA’s software guru Dwight Diercks has promised me to dedicate a lot of time for including 3DNow! into the drivers, 0172b is only the first step into that direction and it improves K6-2 and K6-3 frame rate results by about 10% already. The TNT2-driver follows Nvidia’s tradition and contains a full OpenGL-ICD.
Problems
Fortunately this will be a pretty short paragraph, since I hardly stumbled about many problems at all. The most annoying problem was some rare lock ups in Quake2 when ‘timedemo’ was at ‘1’. I was assured that this is only a driver issue and Dwight is already taking care of it. The next thing I stumbled across was a problem with Socket7-systems based on the Ali Aladdin V chipset. I could only start Windows 98 when disabling ‘AGP Turbo Mode’ in the BIOS setup of the Asus P5A, but once that’s done TNT2 runs without any problems. I’d still prefer to see this issue sorted out though, because the P5A is simply faster with ‘AGP Turbo Mode’ enabled. Last but not least was I quite surprised to see that the latest TNT2-driver does not support the resolution 1280×960 of Quake2 in full screen mode. TNT2 is definitely fast enough to make this resolution pretty useful in Quake2 and NVIDIA has recognized the problem. The release drivers of TNT2 will support 1280×960 full screen mode as well.
Image Quality
TNT2’s image quality is identical to the immaculate image quality of TNT. There is certainly no reason to show you any screen shots, especially not if you are already the owner of a TNT based graphics card. People who are used to Voodoo2’s image quality should really have a look at e.g. Quake2 running on TNT. Many of the Voodoo2-owners seem to still not know how much better Quake2 can look. TNT2 will combine TNT’s image quality with 3D performance way ahead of Voodoo2.
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 from February 10, 1999 |
3Dfx Voodoo2 |
4.11.01.0350-2.18 from December 23, 1998 |
ATI Rage128 |
4.11.6069 from March 3, 1999 |
NVIDIA RIVA TNT2 |
4.10.01.0170 from March 6, 1999 |
NVIDIA RIVA TNT |
4.10.01.0110 from February 1, 1999 |
2D Performance
I tested TNT2 with Winstone99 just to find out that again the results were identical to the other tested graphics cards. None of the tested cards showed any performance decrease in Winstone when switching the desktop color depth from 16 to 32 bit. I recently received a mail from a 3Dlabs-employee,who told me that Winstone would be pretty useless for the evaluation of 2D-performance. Under Winstone the graphics driver is waiting for the CPU to send its next instruction or data, which is why all cards a performing the same as soon as the graphics driver is faster than the rest of the system. He suggested that I run Winbench instead. You can see that I still haven’t done that. The reason for this is pretty simple. If the hectic test run of Winstone is already too slow for today’s 2D drivers, what about us actually using office applications? We are certainly slower than Winstone, so that a faster 2D graphics driver has got even less impact on our real world computing experience. Thus it seems pretty pointless to test 2D-performance with Winbench, because the 2D-performance of a card may be a great as it wants, it doesn’t have the slightest impact on real world computing. My conclusion about TNT2’s 2D performance is: It’s damn fast, but you won’t notice the slightest difference to Voodoo3, Rage128 or TNT.
32 Bit Rendering Performance
You certainly can still remember, NVIDIA does actually not despise the idea of 32-bit rendering and TNT2 does 32-bit rendering, whilst 3Dfx’s Voodoo3 doesn’t. The question is now if 3Dfx is right, is 32 bit rendering a frame rate killer indeed and thus pointless?
I only used Quake2 for the comparison between 16 and 32 bit rendering, because other games that support 32 bit rendering are either too old like e.g. ‘Incoming’ or not complex enough and thus scoring equally in both color depths as e.g. Turok2.
The impact of 32 bit rendering in Quake2 lies somewhere between 1 and 17%, Quake2 is still playable at 32 bit as well in the vast majority of resolutions. NVIDIA must have done a pretty good job with the rendering pipeline as well as the memory interface, TNT was not as great at 32-bit rendering as TNT2 is.
TNT2 at Different Clock Frequencies – Quake2 Crusher
Brett ‘Three Fingers’ Jacobs’ Crusher Demo is my favorite Quake2 benchmark, because it puts the Quake2-performance of a 3D-card into a worst case scenario. This means that the scores in this benchmark are pretty close to the lowest frame rates you should occur when playing online and as long as the lowest scores are high enough, you don’t have any reasons to worry. The high scores produced by the Demo1 or even Massive1-demos don’t really give you a good idea about the frame rate you can expect when playing a heavy multi player session. Crusher is definitely CPU-bound, which is why you cannot reach more than 55 fps at any resolution with a Pentium III 500. Whoever reports higher results with any 3D card is not telling you the truth.
Resolution |
Nvidia Riva TNT-2 175/200 32 MB |
Nvidia Riva TNT-2 150/183 32 MB |
Nvidia Riva TNT-2 125/150 16 MB |
640×480 |
53 |
52.6 |
51.1 |
800×600 |
52.4 |
51.5 |
51 |
960×720 |
51.4 |
50.3 |
48.4 |
1024×768 |
51.2 |
49.5 |
47.3 |
1152×864 |
45.7 |
43.4 |
38.9 |
1280×960 |
Only in window |
Only in window |
Only in window |
1600×1200 |
29.7 |
26.5 |
21.9 |
It is amazing to see that TNT2 is pretty much CPU limited in Crusher up to a resolution of 1024×768. This means that in those resolutions TNT2 will score even higher with a faster CPU. 1152×864 is where finally the 3D-chip limitation kicks in, but look at those numbers! TNT2 scores almost 46 fps at 1152×864! It’s a shame that the current driver doesn’t support 1280×960 in full screen mode, because I am sure that TNT2 at 175/200 will still score close to 40 fps in Crusher, which makes this resolution definitely playable. A Crusher result of almost 30 fps at 1600×1200 shows that even this resolution is still almost playable.
Underneath you’ll find the scores of Voodoo3:
Resolution |
3Dfx Voodoo3 3500 |
3Dfx Voodoo3 3000 |
3Dfx Voodoo3 2000 |
640×480 |
52.5 |
52.1 |
51.8 |
800×600 |
51.9 |
51.9 |
49.5 |
960×720 |
48.2 |
46.5 |
42.3 |
1024×768 |
46.4 |
43.4 |
39.2 |
1152×864 |
40.6 |
37.4 |
32.4 |
1280×960 |
35.2 |
33 |
27.8 |
1600×1200 |
25.1 |
22.6 |
18.6 |
Voodoo3 is already starting to dive at 960×720 and at 1152×864 it’s 5 fps behind TNT2. The next chart makes comparing the two chips a lot easier:
This chart shows pretty obvious who’s the real king in Quake2 now. You can easily see that TNT2 is always faster than Voodoo3. What you cannot see on this chart though is that even TNT2 125/150 is scoring almost as well as Voodoo3 3500, and that’s maybe the most remarkable thing.
TNT2 at Different Clock Frequencies – Shogo Tomsdemo
Tomsdemo for Shogo is currently one of the best real world benchmarks for Direct3D-games. Shogo may not be quite as cool as Halflife or Unreal, but at least it’s graphic engine works without any flaws and the benchmarking procedure isn’t quite as painful and depressing as with the two mentioned games. Shogo is a DirectX6-game, it is using multi-texturing and can even support bump mapping. All in all it’s got a pretty nice and sophisticated engine.
I used the following settings in Shogo:
- High detail
- Optimized surfaces enabled
- Triple buffering enabled
- Sound, music and movies disabled
- Joystick disabled
I found that enabling ‘optimized surfaces’ speeds up TNT, TNT2 and Rage128 by a significant amount. It doesn’t do anything for 3Dfx-cards though.
Resolution |
Nvidia Riva TNT-2 175/200 32 MB |
Nvidia Riva TNT-2 150/183 32 MB |
Nvidia Riva TNT-2 125/150 16 MB |
640×480 |
83.5 |
82.9 |
77.5 |
800×600 |
77.1 |
76.4 |
72.6 |
1024×768 |
69.1 |
65.7 |
56.4 |
1152×864 |
56.1 |
52.5 |
43.4 |
1280×1024 |
46.5 |
43 |
35.2 |
1600×1200 |
30.8 |
27.9 |
22.7 |
I guess 46.5 fps at 1280×1024 resolution is pretty cool really.
Shogo Tomsdemo Frame Rate [fps] |
3Dfx Voodoo3 3500 |
3Dfx Voodoo3 3000 |
3Dfx Voodoo3 2000 |
640×480 |
82.4 |
81 |
75.8 |
800×600 |
71 |
65.9 |
57.9 |
1024×768 |
49.3 |
43.4 |
37.9 |
1152×864 |
39.4 |
35.5 |
30.6 |
1280×1024 |
30.4 |
27.5 |
24.5 |
1600×1200 |
Fail |
Fail |
Fail |
With Voodoo3 you can’t really play Shogo at a resolution of 1280×1024, 30 fps are also completely against 3Dfx’s policy, claiming that anything less than 60 fps sucks badly. I guess this means that Voodoo3-owners can play Shogo only at 800×600 then, whilst a TNT2-owner can nicely enjoy the better detail of 1024×768. The Voodoo3 does not run Shogo at 1600×1200 correctly, all moving object are lacking textures. This is either just a bug in the driver or it’s due to Voodoo3’s lack of AGP-texturing support.
Here the direct comparison chart:
TNT2 kicks the butt of Voodoo3 really badly in Shogo. TNT2 is really strong at Direct3D.
Comparison of Voodoo3 with Actual 3D-Cards – Quake2 Crusher
Now it’s time to take the focus off the fight between TNT2 and Voodoo3 and look at the other 3D cards out there running in systems with four different CPUs.
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 |
39.2 |
34.3 |
36.0 |
32.6 |
3Dfx Voodoo2 SLI |
43.5 |
38.2 |
42.3 |
38.2 |
NVIDIA RIVA TNT |
30.9 |
30.4 |
26.4 |
23.1 |
ATI Rage 128 |
33 |
29.4 |
26.1 |
20 |
Nvidia Riva TNT-2 175/200 32 MB |
51.2 |
39.1 |
29.5 |
23.5 |
Nvidia Riva TNT-2 150/183 32 MB |
49.5 |
39.1 |
29.5 |
23.5 |
Nvidia Riva TNT-2 125/150 16 MB |
47.3 |
38.5 |
29.3 |
23.3 |
As long as an Intel CPU is being used, TNT2 is ahead of the whole competition. However, it’s also interesting to see that the different clock speeds of TNT2 do only make a difference in Q2-Crusher when your CPU is faster than a Celeron 400. Until TNT2’s driver doesn’t get some really good 3DNow!-enhancements, 3Dfx will stay the 3D-chip maker for the owners Socket7-systems. The TNT2-driver 0172b is already a step into the right direction, it makes the TNT2 perform pretty much identical to a Voodoo2-SLI with AMD K6-2 and K6-3.
Comparison of Voodoo3 with Actual 3D-Cards – Shogo Tomsdemo
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 |
37.9 |
36.8 |
37.1 |
34.1 |
3Dfx Voodoo2 SLI |
42.8 |
40.9 |
39.7 |
36.9 |
NVIDIA RIVA TNT |
34.6 |
33.4 |
31 |
29.9 |
ATI Rage 128 |
36.7 |
36.7 |
36 |
34.4 |
Nvidia Riva TNT-2 175/200 32 MB |
69.1 |
56.4 |
40.4 |
33.2 |
Nvidia Riva TNT-2 150/183 32 MB |
65.7 |
54.8 |
40.3 |
33.2 |
Nvidia Riva TNT-2 125/150 16 MB |
56.4 |
51.3 |
39.3 |
32.9 |
In Shogo TNT2 is way ahead of its competitors on Intel platforms, but you can also see its CPU-dependency. The driver 0172b lets TNT2 score the same as Voodoo3 3500 on a K6-3-system, as you will see once I’ve ran the benchmarks again and updated the scores. Interestingly ATI’s Rage128 scores almost identically to Voodoo3 2000, which I find quite remarkable.
Please note also that TNT2 at 175/200 is exactly double as fast in Shogo as TNT! It doesn’t happen very often that the new product of a company shows a performance increase of 100% over the previous product!
AGP-Performance Test with S3’s Quake2 Mon2-Demo
I still really like the ‘mon2’-demo for Quake2 from S3. It’s such a nice way of showing mediocre AGP-texturing performance. 3Dfx hates this demo, because it has always made their products look bad, due to its lack of AGP-texturing support. ‘Mon2’ is a demo running in a custom Q2-level using a large amount of textures.
Traditionally TNT as well as the new TNT2 look very good in this benchmark, due to their proper support of AGP-texturing. Voodoo3 scores pretty bad, but the biggest surprise is the Rage128. ATI has still not fixed the driver, the AGP-performance under OpenGL seems to be non-existent in ATI’s Rage128.
Conclusion
Before I let you know my thoughts about TNT2, let’s try and compare TNT2 to its toughest competitor Voodoo3:
TNT2 | Voodoo3 |
Support of 32-bit rendering, which seems to have an impact of around 10-15%. | No support of 32-bit rendering. |
Support of AGP-texturing, enabling the proper and fast display of scenes with highly detailed textures. | No support of AGP-texturing. |
AGP 4x support | AGP 4x support only in Voodoo3 4000 coming up later this year, but still no support of the AGP-texturing |
Highest frame rates in 3D-games | High frame rate in 3D-games, but less than TNT2 |
Full OpenGL-ICD | OpenGL-ICD still not finished, questionable if it will be available once Voodoo3 ships |
Digital flat panel output integrated into chip | Digital flat panel output on extra-chip called LCDfx, possibly more flexible solution |
RAMDAC 300 MHz | RAMDAC 350 MHz, supporting higher refresh rates at highest resolutions |
32 MB onboard memory | 16 MB onboard memory, thus only limited color depths at highest resolutions |
Available from several different graphics card vendors, thus possibly better prices due to competition between them | Available only from 3Dfx via STB, no competition, prices will only be influenced by products with different chips |
No Glide support | Glide support, possible advantage |
No impressive 3DNow!-support | Excellent 3DNow!-support |
NHVIDIA expects good availability of TNT2 at 150+ MHz | Probably low yield of Voodoo3 3500 |
Expected mass shipment starting mid April 1999 | Mass shipment expect to start end of March 1999 |
I guess the frame rate numbers as well as the comparison above shows it pretty clearly. TNT2 seems to be the better product indeed. It has certainly got the better technology, supporting e.g. 32-bit rendering as well as AGP-texturing. The frame rates scored by TNT2 speak their own language, the 300 MHz RAMDAC should hardly be a problem for the vast majority of us, who had to live with 250 MHz RAMDACs so far. We also will still have to wait what’s happening on the flat panel market in terms of screens with a digital input. NVIDIA offers a full OpenGL-ICD, whilst 3Dfx is working on theirs for a really long time now, alas with no result yet. The most interesting story is hidden behind 3Dfx’s acquisition of STB though. 3Dfx has pissed off a lot of 3D-card makers after buying STB and logically not supplying any other 3D-card maker with their chips anymore. Companies like Diamond and Creative, but also Elsa, Guillemot and even Asus felt as if 3Dfx was trying to put them out of the 3D-business. Those companies were waiting for their chance to kick 3Dfx’s butt and thus they all teamed up with NVIDIA. Now the disaster has happened to 3Dfx. NVIDIA has indeed produced a chip that’s superior to their new and not even yet released Voodoo3 and Creative, Diamond and the likes can’t wait to screw up the sales of Voodoo3. NVIDIA is in a lucky situation, almost everybody in the business is on their side against 3Dfx. I wonder what is going to happen to poor and soon ‘bullied’ 3Dfx in the next few months.
NVIDIA called their new product ‘TNT2’, clearly marking that this new chip is based upon its predecessor ‘TNT’. Still the performance of that ‘TNT2’ lies at magnificent 70 to 100% above the performance of TNT. At the same time 3Dfx created a new chip that is only some 5-15% faster than their previous top-performer ‘Voodoo2’. Instead of honestly calling this chip ‘Banshee2’, which would come close to the truth, they call it ‘Voodoo3’, suggesting a new technology. It seems pretty obvious that NVIDIA overtook 3Dfx only by continuously doing their homework properly. The mediocre performance increase of Voodoo3 over previous 3Dfx-products as well as Voodoo3’s lack of new technologies like e.g. AGP-texturing or 32-bit rendering make it look as if 3Dfx’s developers either got lazy or they are simply not up to doing their job right anymore. It may sound harsh, but the time was due for 3Dfx to lose their leading position. The Romans used to say ‘Carpe diem!’ – NVIDIA did, 3Dfx didn’t.
From what I have seen of Voodoo3 and TNT2 I would not even hesitate a second and go for TNT2. It’s overall the better chip and it will most likely even be cheaper than Voodoo3. As long as NVIDIA hasn’t finished the 3DNow!-optimized driver though, AMD K6-2 or K6-3 owners may rather prefer Voodoo3, because it has an excellent 3DNow!-support.
NVIDIA has in any case got all reason and all rights to throw a huge party now. They’ve managed to push 3Dfx from the throne that 3Dfx occupied for over 3 years. The company that produces the fastest 3D-chip for PCs is from now on to be called NVIDIA! What have I read recently? “The King is dead, long live the King?” – I may be excused for altering that a bit to “The old King lies on his death bed, long live the new King from a different and more progressive tribe”. Let’s hope that the new king won’t get as lazy as the old king did …