The Wind of Change
Times are changing in the graphics business. Prices of graphics cards are dropping and this might change a lot in the card makers business. We shouldn’t be surprised if the company structure of well renowned businesses as Diamond or STB should look a lot different in twelve months from now. The trend is going towards cheap card productions and designs from Taiwan or even better the combined chip and card production by the same company, as e.g. very successfully done by ATI. The price awareness of the people is a lot higher than 12 months ago and why should anyone be surprised about it? Who likes to spend a lot of money for a product that’s obsolete in 3-6 months?
Another change had to be learned the hard way by the former technology leader in the PC 3D graphics sector. 3Dfx’s sales are dropping and the sales of boards with Voodoo2 chip are “below expectations”, e.g. Diamond has got a large stock of Monster 3D II boards because they expected to sell a lot more of then than they do. The latest product of 3Dfx, Voodoo Banshee, is certainly no bad product at all, however it can not satisfy the expectations people had of the ‘technology leader’. Banshee is good, but not good enough, more importantly it is not better than the previous, although different, product Voodoo2.
As if this wasn’t bad enough, now there is a chip out that is even up to threatening Voodoo2 itself. 3Dfx is facing tough times after 2 great years, they need a new product that will rock the boat pretty soon.
The Voodoo2 challenger is NVIDIA’s new RIVA TNT chip. Now NVIDIA didn’t and still doesn’t have an exactly great time themselves. First of all the product specs published in the first press release about TNT had to be retracted and replaced by lower specs. This cost quite a bit of reputation and the people became pretty skeptical. Now since it looks as if TNT was still a great product, there comes 3Dfx and files a lawsuit against a business that is trying its hardest to go public and get some cash. Not that NVIDIA wouldn’t be experienced with lawsuits. No, almost every important company in the 3D area filed lawsuits against NVIDIA. However, you don’t really need pending lawsuits just before your initial public offering. The lawsuit filed by 3Dfx is of some touchy nature. It’s about 3Dfx’s patent on ‘multi-texturing’. I don’t want to rant about the fact that it’s included into DirectX6 and that ATI as well as 3DLabs are implementing it into their upcoming chips as well, but I wonder why the very company owning the patent for ‘multi-texturing’ is the one that is NOT using it in their very own 2D/3D solution called Banshee.
Anyway, I do certainly realize that there are more competitors of Voodoo2 out there than only NVIDIA, but the RIVA TNT is available and has the highest potential of competing against the current 3Dfx products. Thus I won’t talk about Matrox G200, because it’s not fast enough, I won’t talk about S3’s Savage3D, because it’s still not stable and bug free enough as well as slower than TNT, I will also not talk about ATI’s Rage 128 nor 3DLabs’ Permedia3, simply because those chips are still closer to the scratch board than reality. Oh yeah, and I haven’t forgotten about Rendition as well, but tell me, is this company still alive?
This article shall answer the question if NVIDIA’s RIVA TNT will be able to replace Voodoo2 and if it’s indeed better than Banshee. I did a whole lot of research about that and I will be able to give you a clear answer. To already answer this question for my part, yes, I am using TNT and there are no 3Dfx cards in my own system anymore, the first time after more than 2 years. Let’s find out if this will be feasible for you too.
3D Benchmarking and What Is Really Important for Us
I had a meeting with an old friend from the graphics business today. We spoke about the good old time when my website was still very young and I came up with the ‘Monster Truck Madness 3D Benchmark’ (Nov. 3, 1996), as some of my long-term readers may still remember. He said that this was the first time that a game was used for benchmarking and when he went to magazines and told them to use it, he was asked by the editors “what? You want to use a game for benchmarking?”
A lot has changed since then. Many games include some way of benchmarking, Quake was and now Quake2 is the most known one. All those game benchmarks have one problem however, they tell us the average frame rate over a certain amount of played frames or after a certain time of demo play. Since the release of Voodoo2 we got pretty used to frame rate numbers in the range of 50 to even over 100.
However everyone knows that the eye needs only 25 fps to see a scene as a smooth movement, movies are played at 24 fps and everyone seems to be happy with it. The need for frame rates way above 25 fps was justified with the theory that in game play the frame rate could drop down to 10 fps for only one second, thus hardly changing the average frame rate reported by a benchmark, but making the game unplayable for this moment. This is why frame rates of 35 fps were rated as not good enough, or why a card that does 70 fps instead of 35 fps is regarded better.
If we think about it for a second, we can see that this way of rating the 3D cards is not quite correct. Couldn’t a card that is scoring an average of 70 fps be running at only 15 fps in some scenes, but equalizing this by running at some crazy 125 fps in another scene? Couldn’t a card that does 35 fps be only varying frame rate between 30 and 40 fps and thus be even better than the 70 fps product? The fact of the matter is that we shouldn’t care less about the wonderfully reported average frame rate.
What we really need to know is the minimal frame rate in a game and if possible we should see how the frame rate changes in game play. Luckily I found a partly solution to that. I’ve now got software that saves the frame rate for each frame of a demo, so that you can analyze it later. For the first time I was able to see a graph that shows how the frame rate changes from frame to frame.
Unfortunately I can only do that with Direct3D games as yet, but I hope that I will be able to do this for OpenGL games soon too.
The Meaning of ‘Flip at Vertical Retrace’ or ‘VSYNC’
Matrox was the first company that included a feature in their good old Mystique, which enabled the switch to a new frame before the previous frame was displayed completely at the CRT screen. This caused a lot of anger within the competitors, because Microsoft’s DirectX specification did and still does not allow this feature, because it can cause ugly tearing. On the other hand it can increase the average frame rate score of a 3D card by a fair amount. Last year there were still a lot of 3D chips which would not include this feature, as e.g. NVIDIA’s RIVA 128, before it became common that everybody was benchmarking 3D cards with games. Today the ‘Disable VSYNC’ feature can be found in the drivers of every 3D chip out there and it’s time to wonder if this is only done for benchmarking or if it can improve performance in a normal environment as well.
When looking at it from a theoretical point of view, it seems clear that nobody would need ‘Disable VSYNC’ for normal game play, because there’s certainly no point of drawing new frames to the screen more often than the monitor refresh rate. The lowest monitor refresh rate today is 60 Hz and the human eye only needs 25 pictures/second to get the impression of a smooth movement, so even the limit to 60 pictures/s (‘VSYNC endabled’) would be way higher than what a human eye requires. This means that in theory people who are using Microsoft approved ‘WHQL’ drivers, which do not include the ‘disable VSYNC’ feature, shouldn’t have to worry about a noticeable performance disadvantage. An editor from the German PC Professionell is even benchmarking with ‘VSYNC enabled’, because he says that this is the way people are using the cards in reality. Of course the frame rate can never exceed the refresh rate in this case, thus lowering the ‘average frame rate’ result.
When I was running my tests with this new little software, I was running some benchmarks with ‘VSYNC enabled’ as well. You would expect the same graph as with ‘VSYNC disabled’, only the peaks clipped at refresh rate. However, the situation is rather different, as the graph below shows. The frame rate is not only clipped at 85 fps, according to the refresh rate of 85 Hz, but also at the next lower frame rate of actually 42.5 fps, which is half the refresh rate. It seems that with ‘VSYNC enabled’ in some games the 3D card can either draw at refresh rate or at half or a third or quarter of the refresh rate only. This would mean that a new frame is only drawn either at every refresh or at every second or third refresh only, not any time in between.
As you can imagine, ‘VSYNC enabled’ can have quite a significant performance impact. It means that with ‘VSYNC enabled’ the card may not ever draw a picture every refresh, but only every second refresh, limiting the frame rate not to the refresh, but to half the refresh rate. A good example is Expendable running at 1024×768 on Banshee. You can see that the frame rate is clipped at 42.5 Hz, although you wouldn’t expect any clipping at all, since the frame rate with ‘VSYNC disabled’ is never reaching 85 fps and should thus stay untouched by the VSYNC issue.
We should come to the following conclusion. Benchmarking should by all means be done with ‘VSYNC disabled’, and it seems also advisable to disable VSYNC even for normal game play.
Benchmarking
This time I tried to use not only some well known benchmarks as Quake 2 and Incoming , but also the already well known game Unreal and Rage’s upcoming DirectX 6 game ‘Expendable’. Quake 2 is still a very widely played game and could not possibly be missing in a benchmark comparison. Besides Quake2 there are several other upcoming games that will use the Quake 2 graphics engine, so that the Quake 2 benchmark will stay very important for quite a while. Unreal is a game that never really quite did it for me, but it’s also pretty widely spread. More importantly there are over 20 licensees of Epic’s Unreal graphics engine, so that the benchmark will also stay important way beyond the life of Unreal itself. It’s still very touchy to benchmark with Unreal, because it was written for 3Dfx’ proprietary Glide engine and Videologic’s PowerSGL engine in the first place, and the port to DirectX 6 is still not finished. Thus I had to run Unreal with its Direct3D beta patch, which is still a lot slower and lot less stable than the final version will be. Nevertheless I thought it would still be interesting to see how TNT scores in Unreal right now.
I was looking around for a real DirectX 6 game and luckily I received the rolling demo of Expendable from Rage software. Expendable is a quite unique game, offering arcade shooting for two people on the same screen. It has formidable effects and is thus very impressive to watch.
This is what it looks like. It does not come with any built in frame counter, but thanks to my testing software I don’t really require that. Expendable is another typical Rage game, pushing the edge of technology and offering a unique way of game play. For benchmarking it’s perfect too.
I would have liked to include a few more DX6 games, but not all of them are useful for benchmarking. Shogo from Monolith will be an absolutely great benchmark, however the downloadable demo that’s currently available doesn’t enable benchmarking yet. Valve’s Half-Life is another game that I would have loved to use for benchmarking. I’ve got the Day One demo and I played through it within a few hours, running it on my own TNT card at 1152×864. It looks stunning, but the atmosphere is what really gets to you. Half-Life has the most gripping atmosphere since the good old ‘System Shock’ from Looking Glass. Both games will be used for benchmarks as soon as I can get my hands on final versions.
Benchmarking Setup
The tests were ran with 3 different systems, a Pentium II 400, a Celeron 300 A and a K6-2 300.
All systems were using:
- Hard Drive: IBM DGVS 09U, ultra wide SCSI
- SCSI adapter: Adaptec 2940 UW
- Memory: 128 MB PC100 SDRAM
- Operating System: Windows 98
- Refresh Rates: 85 Hz for all cards in all resolutions except 1152×864 and 1600×1200 75 Hz
- Quake 2 ver. 3.19
- Unreal original release version, plus D3D patch for TNT, ver. 209 for Banshee
Motherboards:
- Asus P2B for both Slot1 CPUs
- Asus P5A for AMD CPU
Drivers:
- NVIDIA RIVA TNT: 4.10.01.0036, Chip clock 90 MHz, Memory Clock 112.5 MHz
- 3Dfx Voodoo2: 4.10.01.0180-2.17
- 3Dfx Voodoo Banshee: 4.10.01.0090-1.00
Quake 2 Benchmark Results
I certainly don’t play many games, simply due to time restrain, but Quake 2 has gotten me and thus I am playing 1-2 hours online each day. I’m also the proud member [F3] T-Bone of the clan F3 (Fanatic Frag Freaks). Thus I reckon myself as pretty experienced about what really counts for a Quake 2 player. I know that the taste for weapons is widely spread amongst the Q2 players and whilst the majority seems to love the Rocket Launcher, I love the Railgun. This weapon is one of the most powerful in experienced hands, but it asks for two important things. A good ping and a high screen resolution. The Railgun is an excellent sniper weapon and at a high resolution you can kill enemies at large distances who may not even see you. You can also aim a lot more smoothly with a higher resolution, so for me screen resolution is the way to go.
I don’t see any reason for playing Quake2 at 640×480 anymore since the days of Voodoo2 and TNT. It is advisable to play at the highest screen resolution you can get, as long as the frame rate is acceptable. Now here the discussion comes in. What is acceptable? Many Q2 players will answer ‘40-50+ fps is what it takes’. This is of course incorrect. A sustained frame rate of 25-30 fps would actually do, it may never ever drop below 25 though, because that would cause game play restrictions. The different fps values that people are talking about come from different benchmarking demos for Q2. There is the pretty useless built in ‘demo1.dm2’ or ‘demo2.dm2’, or some more demos included into ‘Reckoning’, the Q2 mission CD. Then there are the two larger demos recorded from the Q2 benchmarking Guru Brett ‘3 Fingers’ Jacobs, which are ‘massive1.dm2’ and ‘crusher.dm2’. You may remember what I said in the introduction, we shouldn’t care less about ‘average frame rate’, we want to know the worst case scenario of a 3D card. As long as this is above a decent level, we don’t have to worry anymore. Now unfortunately Q2 tells us only the average frame rate from a benchmarking demo, no maximum, no minimum. Luckily did Brett help us out here. He had the idea of putting 3D cards through tougher tests by recording a demo with a huge amount of explosions and light weapons, so he recorded ‘crusher.dm2’. I never used this demo so far, because in reality it’s a perfect CPU benchmark. The CPU rather than the 3D card do the transform and lighting and ‘crusher’ is asking for the last bit of CPU performance. This is what makes it so great though. ‘Crusher’ can be seen as the worst case scenario in Quake2. You will hardly ever be in a death match with so many rockets, hyperblaster , BFG and Railgun shots flying around you, or if you should, you may not stay alive for long. Using ‘crusher.dm2’ shows how bad the frame rate can get. I claim that you can run Q2 just fine as long as your ‘crusher.dm2’ result is above 25 fps. Now what we really want to know is how far can we go up with the resolution whilst keeping a good result in crusher.dm2.
Running Quake 2 3.19 with 3Dnow! Enabled
Owners of AMD’s K6-2 CPUs are always a bit in trouble if they want to play Q2 online. 99% of the available Q2 servers are running the latest Q2 patches as soon as they become available. The 3Dnow! Patches for those Q2 upgrades can often take 4-6 weeks and so no K6-2 owner can play online Q2 in this time, which really sucks.
I tried finding a way around that, because I wanted to benchmark all CPUs with the same Q2 version. To run Q2 3.19 with 3Dnow! enabled you don’t have to do much at all. Simply copy the file 3dfxglam.dll as 3dfxgl.dll and ref_glam.dll as ref_gl.dll into the Quake 2 base directory, replacing the original drivers (rename them to keep them). Now 3Dnow! is enabled each time you either run a 3Dfx card with the 3Dfx OpenGL driver or e.g. TNT with the default OpenGL driver. There is one little problem though. The ref_glam.dll would not let you disable dynamic lighting (gl_dynamic 0). It just ignores this command and you won’t get the performance increase unless you enable flash blend (gl_flashblend 1), which for some reason does finally switch off dynamic lighting in the default OpenGL 3DNow! driver ver. 3.17. You should rather use the original ref_gl.dll with K6-2 in case you want to benefit from the performance increase achieved by disabling dynamic lighting and don’t want to have the stupid effect of flash blend.
Quake 2 Results with NVIDIA RIVA TNT
With Intel’s Pentium II at 400 MHz and TNT you can play Quake 2 just fine at 1152×864, a resolution that is not achievable on any current 3Dfx product. It’s pretty pointless to look at the results at 640×480 or 800×600 in my eyes. Why running at those low resolutions if it runs perfectly at higher resolutions?
The same is pretty much valid for Celeron 300 A or Pentium II 300 users. 24 fps in crusher are still good enough for decent game play at 1152×864.
Now 18-19 fps in crusher are a bit touchy, so you may experience some (maybe only minor) slowdowns when playing Q2 in some occasions. However please note that it makes hardly any difference if you play at 640×480 or at 1152×864, thus I’d go for 1152×864 in the first place. If you try and tweak Q2 only a little, you get a lot better results.
I used only those tweaks which you can either live without just fine or which produce some real performance increase.
- Switching off dynamic lighting has the largest impact of about 5-6 fps. Dynamic lighting lightens up the area around a flying object like a rocket or a blaster shot, and to be honest, although it looks nice, you can really do without it if performance it at stake. You achieve that with the command ‘gl_dynamic 0’. You can either put this line into your autoexec.cfg, or into a special cfg file, or you can type it in the console.
- A real must for every serious Q2 player is switching off ‘polyblend’. This feature is responsible for the most annoying flash of the screen when you collect an item (like a weapon, armor or ammunition) and is also making you half blind if you have quad damage or invulnerability. It also impacts the visibility under water. All in all it’s just this one thing you don’t need. It’s not nice, but it’s definitely annoying. You need the command ‘gl_polyblend 0’ and you got rid of it. It gives you about 0.5-1 fps as reward as well.
- It is more controversial to actually switch off the weapon particles. This command lets you get rid of the little white dots you see when e.g. firing a rocket. This is the good part of it, because I can do without those blinding particles in rocket fights just fine. However, it also makes the nice blue trail of the railgun go away, which always shows you where a railgun shot came from as well as how much and where you missed your opponent. For me, the particles have to stay for this reason, but people who don’t care about the railgun could do just fine without it. The command is ‘cl_particles 0’. It buys you about 1-1.5 fps.
The following results are achieved by running Q2 with all those three features switched off:
- gl_dynamic 0
- gl_polyblend 0
- cl_particles 0
It is surprising to see that with those tweaks, you get 25 fps in crusher.dm2 with a K6-2 300 and TNT at 1152×864!! Thus it is not correct to say that TNT is not doing well enough with slower CPUs. There is no other graphics card available that would offer you playable Q2 at 1152×864 resolution with a K6-2 300! Please note that the ref_glam.dll would not let you disable dynamic lighting, which is why it scores worse than the default ref_gl.dll after the tweak.
Quake 2 Results with 3Dfx Voodoo2
There is not much to say about the results. Voodoo2 is running Q2 just fine with all CPUs at all offered resolutions. One Voodoo2 card can’t do more than 800×600 and even two Voodoo2 cards can’t offer more than 1024×768 resolution however This is the big disadvantage of Voodoo2, the ‘pixel per dollar’ ratio is not very good.
Quake2 Results with 3Dfx Voodoo Banshee
You can see that Quake 2 with a Pentium II 400 and Banshee at 1024×768 is already too slow for reliable online game play, but with some tweaking you could get it running properly.
A Celeron 300 A or Pentium II 300 let’s you play Quake 2 with Banshee at 800×600 only, unless you tweak Quake 2.
These results are not quite representative, because the 3DNow! driver for Quake2 wouldn’t run with Banshee. I don’t know if we can expect the same results as for a Pentium II 300, but again you see that 1024×768 is no real option without tweaking.
All in all is Banshee not the right 3D chip for people who want to play Quake 2 at higher resolutions, and it doesn’t go beyond 1024×768 as well.
Summary Quake 2 Results
If you want to play Quake 2 at 1152×864 you’ll need TNT, that’s simply it. Even K6-2 owners can get it to run at this resolution in case they don’t mind tweaking Quake 2 a bit. Voodoo2 is still great for Quake 2 and if you run 2 Voodoo2 cards with SLI you’ll get absolutely safe frame rates for each of the tested CPUs. However, 1152×864 is simply impossible. Another thing to look at is image quality. Please take my word for it and forgive me the lack of screen shots, but TNT looks a fair amount better than Voodoo2 or Banshee does.
All in all is Quake 2 certainly no reason to NOT buy a TNT card. For me it would be the reason to buy one, but I really fancy the 1152×864 resolution and you may not. Take in consideration that the two Voodoo2 boards cost you at least double the money that TNT costs and then you still need a 2D card to go with Voodoo2. In my eyes TNT already won the race as the best Quake 2 graphics card. Screen resolution plus image quality made me fall in love with it.
Unreal Benchmark Results
Unreal is out now for quite a while and still it doesn’t officially run on much more than Glide and PowerSGL. There’s an OpenGL port for Verite 2×00 cards available now as well, but it’s still in beta state. Recently a Direct3D patch popped up, again only beta and unable to run with any of the Unreal patches, but there’s at least some way of comparing all the different 3 cards with each other. To me Unreal seems to be in beta state altogether still. I’ve hardly come across a game with as many crashes and bug reports, incompatibilities and apologies from the developer for unperfected software. The game itself is pretty ok in single player mode, although it can’t compete with the atmosphere and story line of Half-Life. In multi-player mode it never made it against the much older Quake 2 and I doubt that it ever will. I can see myself playing Half-Life multiplayer rather than ever Unreal, but I guess that’s a question of taste.
Nevertheless is the engine of Unreal great to look at, the effects are rather impressive and Unreal is supposed to have over 20 licensees for this engine. That’s why Microsoft is working hard on a Direct3D port of the Unreal engine and I hear from NVIDIA that once it is released, TNT will “at least have Voodoo2 performance in Unreal, but look a lot better”. This is all future but now is now, so that I’ve dared to compare TNT running the D3D beta vs. Voodoo2 and Banshee running Glide. Banshee does for some strange reason only run with patch 209, the original release wouldn’t work with it, neither patch 216 or 217.
TNT has quite a problem with Unreal right now still. Decent game play is possible at 640×480 and that’s it. Any higher resolution makes Unreal run that sluggish, that it’s really no fun to play.
Voodoo2 is running Unreal with its Glide engine and this gives it an important advantage. Glide has a lot less CPU overhead and thus it’s possible to play Unreal pretty much at all the available resolutions.
Banshee seems to have a problem with Unreal still, only patch 209 worked. However, you can play Unreal fine at 800×600, which gives it an advantage over TNT. Again, the Glide engine is responsible for this.
Summary Unreal Results
As long as 3Dfx can play its Glide muscle, it can show its superiority over other cards. Direct3D is used much more widely, but it’s also a Microsoft product and Microsoft products are not exactly famous for being modest in its demands for system performance. Direct3D needs a lot more CPU power than Glide and thus the clear winner in Unreal is currently still 3Dfx. This may change in the next 6 weeks however, let’s see what the Microsoft and Epic guys will have achieved by then.
Expendable Benchmark Results
Expendable is one of the new games Rage will launch by the end of this year. It is a DirectX 6 game, which means that it’s taking advantage of the new features of DirectX 6, one of them is e.g. multi-texturing. Expendable is still at a pretty early stage, but it’s running wonderfully. You will see that Expendable is requiring a whole lot of CPU power, so that the CPU and not the graphics card is the delimiter of the frame rate. You will see that the 3Dfx products don’t look as shiny when playing in the same ballpark as TNT, which is Direct3D.
You can see that it’s currently still quite tough running Expendable on a K6-2 300, even at a resolution of only 640×480. TNT seems to have the least performance loss in the driver, making it the fastest chip in this comparison.
At 1024×768 there is a strange occurrence with TNT. It seems to hit its fill rate maximum when driven by a Pentium II 400. This is the only explanation why it’s suddenly slower than Voodoo2 or Banshee, since it’s faster with the other two CPUs. We shouldn’t forget that TNT has got the lowest pixel fill rate of the three, 90 Mpixels/s, Voodoo2 SLI has got 180 Mpixels/s and Banshee has got 100 Mpixels/s, which could be the reason for the lower performance.
Now lets have a look at the lowest frame rates as well.
Again, TNT seems to have the least driver overhead and is thus scoring best at 640×480.
The lowest frame rate shows a different picture than the average frame rate. TNT is here better than Banshee. However, I’m not quite clear why Voodoo2 SLI is faster than TNT, in case of the lowest frame rate it cannot be the fill rate.
Expendable Frame Rate Diagrams
The following graphs shall give you a better idea about what happens when a game is playing and how the frame rate is changing.
You can see that the shape of the graph is always the same, but amplitude and offset change from CPU to CPU.
As I discussed before, there is something strange with the run of the Pentium II 400. You can see that the purple graph is a lot more fuzzy and missing some of the peaks. Something around 50 fps seems to clip the frame rate. I wonder if NVIDIA will have an explanation for it.
The graphs of the 3Dfx chips look all the same, so that I’ll show you the ones of Voodoo2 SLI.
The graph of Voodoo2 SLI for 640×480 is hardly any different to the graph of TNT at the same resolution. This is because the CPU is limiting the frame rate.
If you have a close look you will see that this graph looks identical to the one above. Voodoo2 SLI does not have any problems at 1024×768 with a Pentium II 400, unlike TNT.
Summary Expendable Results
The 3Dfx products have hardly got any advantage over TNT under Direct3D as long as the game is CPU intensive. In some cases TNT is even faster than the much more expensive double Voodoo2 configuration. Voodoo2 SLI does still have a slight edge over TNT at high resolutions, but this can hardly justify the price difference between the two.
Incoming Benchmark Results
Incoming is a well-known DirectX 5 game with a lot of fancy graphical features. It’s also not exactly independent from the CPU performance, but still a lot less hungry than Expendable. You will see that Incoming allows the graphic chip to be the delimiter of the frame rate, at least when using a fast CPU at high resolutions.
At 640×480 it’s obviously not the fill rate that limits the frame rate, because otherwise TNT wouldn’t be as fast as Voodoo2 SLI. It’s still interesting however, that TNT is the fastest 3D chip for the K6-2 in this comparison.
The story looks different at 1024×768. The huge fill rate of Voodoo2 SLI is blowing the competitors away. TNT is obviously at it’s fill rate limit, regardless which CPU is used. The K6-2 is even faster than the Celeron 300 A, which shows that the TNT driver must have some 3Dnow! optimization. It’s surprising to see that Banshee isn’t faster than TNT, although it has the higher fill rate.
The much more important lowest frame rate shows an edge of TNT at 640×480. It also shows that Banshee is a fair amount faster than a single Voodoo2 card. Incoming doesn’t use multi-texturing, so that the second texture unit is completely unimportant in this game, making single Voodoo2 fall behind.
Again Voodoo2 in SLI is by far the most powerful 3D chip at 1024×768 in a game that doesn’t push the CPU too hard. As already mentioned, it’s a pure fill rate issue here and Voodoo2’s brute force makes it win this comparison.
Incoming Frame Rate Diagrams
You will see that things can look a lot different to what you’ve seen with Expendable.
At 640×480 the picture is similar to what we saw in Expendable. The graphs are all of the same shape, but the amplitudes are shifted. TNT is scaling with CPU performance.
The picture changes completely at 1024×768. Now the graphics chip is the limiting factor, all CPUs are showing the exact same frame rate.
Again Voodoo2 SLI looks almost identical to TNT at 640×480.
The story looks different at 1024×768. The second half of the demo run is still able to scale with the CPU. If you remember the demo then you’ll know that the first half is the cannon scene, where all the buildings have to be drawn all of the time, the second half is the helicopter scene, which is a lot less complex. The first half seems to push even Voodoo2 SLI to the limit, the CPUs seem to perform almost identical. The second less complex half can still show a difference between the three CPUs.
Summary Incoming Results
A game that does not require quite as much CPU power is still able to give the double Voodoo2/SLI configuration an advantage over TNT. However, we don’t benefit from frame rates over 30 fps whilst we do benefit from higher resolutions. It doesn’t help that Voodoo2 can reach astronomical frame rates at 1024×768, if it cannot produce any higher resolutions.
Overall Summary
NVIDIA’s RIVA TNT is not the new wonder chip as some people may have expected. However it is sticking up very well against its toughest competitors from 3Dfx. 3Dfx has still got an edge in applications that are available in a Glide version and in games that don’t strain the CPU as much, thus giving a dual Voodoo2 configuration the chance to show its power. However, there are many occasions where TNT is at least as good as single Voodoo2, dual Voodoo2 and certainly better than Voodoo Banshee. At the same time is TNT offering excellent frame rates and excellent image quality at resolutions that none of its competitors from 3Dfx are able to offer. So it may be true that Unreal is currently best played on a Voodoo2 in SLI and that there are a few more Glide applications that make the 3Dfx products look better, but in the long run TNT will easily win. The clear advantages
- price,
- higher screen resolutions,
- image quality significantly better,
- AGP 2x texturing,
are more important than huge frame rates of Voodoo2 SLI in some applications.
Of course I also found some problems of TNT, but those are minor driver bugs or only significant to a minority. The one problem I have is that I get scrambled stuff after switching back from the Windows desktop to Quake 2 in full screen mode, but that isn’t done by many people and NVIDIA is already working on a driver fix. The second problem is the huge amount of current that TNT requires. Some LX motherboards are unable to supply this high current, which could make it impossible to run TNT on those motherboards. One example is the Asus P2L97.
I enjoyed playing Half-Life Day One on TNT at 1152×864 and I’m playing Quake 2 online at this resolution on TNT every day for two weeks now. Shogo is looking great too on TNT, and it also runs fine and smooth at 1152×864. I wonder what should make me spend a whole lot more money on two Voodoo2 cards, which cannot supply the high screen resolution and which take a lot of space in my system as well as producing a lot of heat. There’s certainly still quite a lot of justification for Voodoo2, but if you are wondering which 2D/3D combo to buy. go for TNT and see if you feel like adding two Voodoo2 cards later on. I doubt that you will.