T&L vs fill rate
The basic premise of this argument is whether more detailed worlds (T&L) or higher resolutions (fill-rate) are better. T&L offers game developers the ability to have better looking worlds by allowing them to use more polygons and dynamic lighting. This is good, because instead of having a character model having a block for a head and the details "painted" on using a texture, the head can be actually be shaped like a head, with a nose, eyebrow ridges etc. Lighting, while not as important, provides for lights that can move around, say like fireflies, and is more of an atmospheric effect. As well, onboard T&L removes load from the CPU, allowing better sound, AI, physics etc. in the game.
Higher fill-rate allows games to be displayed at higher resolutions at acceptable frame rates, say 60 frames per second (fps). As well, many effects can be performed with high fill-rate, including bump-mapping, anti-aliasing and higher-level filtering (trilinear, ansiotropic) among others. In general, you want a high resolution to make the pixels less visible, and to see more detail. Bump-mapping makes flat polygons appear to have bumps on them, and anti-aliasing further removes visible pixels, or jaggies, by rendering several pixels displaced less than a pixel width apart and averaging them. Filtering is comparing a pixel with the ones around it; higher-level filtering compares one pixel to more pixels than low-level filtering (bilinear). All of these improve the visual quality of the image.
In an ideal world you would have a card with high fill-rate and T&L. Currently, that's not available. So which is better? Well, right now, not a huge number of games support T&L, so if you have a fast CPU, then a decent fill-rate is more important, however, T&L is still beneficial in terms of freeing up the CPU for sound, AI etc. For people with slow CPUs, say less than 400MHz, T&L offers the chance to increase their frame rates by removing the CPU as the bottleneck.
In the future, games are going to get more complex in terms of polygons, so T&L will become of increasing importance. Fill-rate demands are also going to rise, although not as rapidly as T&L, due to the fact that other parts of the hardware (memory, bus speed) will become bottlenecks to frame rate rather than the actual chip. Actual fill-rate will increase as the chips move to smaller fabs, and this should be enough to keep decent frame rates.
Alright, so that was a pretty crappy comparison that didn't actually conclude much, but it will make more sense when I get onto the actual cards.
3D Cards
Enough theoretical, let's see what is actually available. The three main up and coming contenders for the 3D card crown are the GeForce, the Savage2000 and 3dfx's new product, which I'll call the Napalm, because it sounds cool :). I'll also briefly comment on Glaze3d, and the G400.
GeForce
The NVida GeForce is the immediate heir to the 3D graphics card crown, with the Napalm delayed, and the Savage2000 not available yet. Despite having 480 MTexels (and 480 MPixels) which is more than any current card on the market (the Voodoo3 3500 has 366, TNT2 Ultras have 300-360), some people (whoever wrote the computer industry sanpshot part 2) have claimed that this is not sufficient for GeForce's onboard GPU, which does 10 million triangles per second (a Pentium III 600 can do ~3 million triangles per second and nothing else (AI, physics etc.)). Personally, I think that 480 MTexels is sufficient for today's hardware, especially given that frame rate is becoming ram speed limited. This is based on the performance increase seen by using DDR (double data rate) ram instead of SDR (single data rate) ram at high resolutions (>800x600). As well, games moving from light maps to vertex lighting, as is now possible, will free up a fair few texels.
This fill-rate will also go up when the GeForce is moved to a .18 micron process (it is currently .22). Unlike the aforementioned computer industry sanpshot part 2, I don't think that this will be a sufficient increase in fill-rate to keep the GeForce competetive with the other cards. IMO, now that NVida has put the hard yards in by getting T&L on-chip, they should not spend time improving it greatly for the next generation, as it has ample capacity now. Instead they should put their effort into getting two texels per pipeline. Simply soldering more pipelines together (which is what they did in the GeForce; put two TNT2 pipelines together), or the increase gained by going to a smaller process is not going to keep up with the fill-rate of other chipsets. Getting another texel will increase their fill-rate to 960 MTexels, as good or better than the next generation of cards. Certainly this will add more transistors, and at 23 million already it's not a small chip, but with the smaller process it should become possible to fit them on the die.
Savage2000
With ~700 MTexels and a GPU, is the Savage2000 the gamer's dream come true? Only time will tell, but on paper it certainly looks good. The main problems that could prevent it from living up to its claims are ram speed and the core clock. Since we've seen the GeForce experiencing frame rate speed increases when using DDR ram, and it only does 480 MTexels, the Savage2000 must support DDR ram and a 256-bit memory bus to fulfill its fill-rate. Without it, it will be limited greatly, and become a low-end card, like the TNT2 M64 (which only has a 64-bit memory bus compared to the normal TNT2's 128).
700 MTexels from two pipelines implies a core speed of 175MHz, which seems like a bit of a pipe-dream to me. I don't know what process this is based on, but if it's a .22 micron process like the GeForce, this seems pretty unobtainable, given that NVida only clocks them at 120MHz (overclockers have reached 140MHz). The only way I can see the Savage2000 reaching these speeds is if its GPU is less complex than the GeForce. Admittedly having only two (albeit a bit more complex) pipelines would increase the core speed, but not to that great an extent.
Napalm
The T-Buffer and full-screen anti-aliasing at 1024x768x60Hz implies that the Napalm will have fill-rate to burn, some people have suggested even 1 GTexel. Certainly the visual quality will be excellent (unless it's still only 16- (or 22-) bit graphics), but only if you have a fast enough CPU to support it. This will remove 3dfx from the low end of the market, as there is no point paying for a super-fast card when your frame rate is CPU-limited, as would be the case for (my estimate) anything less than 450MHz. This speed will only increase in the future, as game makers use more polygons to take advantage of T&L.
So, 3dfx will retain its lead in the fill-rate stakes, but only for the high-end gamers who have the money to support it with a fast CPU. As well, it will almost certainly require fast DDR ram to sustain these frame-rates, which will further add to the price. If 3dfx have moved beyond rehashing their one-pipeline Voodoo series and added another pipeline or three, or even come up with a new format, it will mean they will have the skill to put T&L onboard. If they don't put T&L on board in one or two generations, 3dfx will almost certainly become a second-rate player. This is because polygon counts are only going to increase, and people are not going to shell out for a new CPU to keep up with them and an ultra-fast 3D card to get high frame rates at high resolutions; they'll compromise on a GPU and save at least half the cost.
G400
Matrox's latest card has several features that set it out from the pack. Dual head and environmental bump mapping (EMBM) provide an improved experience for the gamer. Dual head is the capacity to connect two monitors to the one card, which allows you to have say a large monitor as your main screen, and another, smaller monitor where you can put WinAmp, ICQ and other stuff you keep open but don't want using up screen real estate. I don't know if dual head is supported by any games, or if it's even possible to use two monitors in a game.
EMBM is the best-looking form of bump-mapping. I explained bump-mapping a bit above, but to give an example, a wall could just be one polygon yet appear to have indentations between the bricks where the mortar is with EMBM. This greatly increases the visual quality without a great decrease in frame rate. However, the success of the G400 will depend on the extent to which these two features are implmented (and its fill-rate as well, but I don't know any numbers). Dual head is probably the easier of the two to implement, but I don't know how many people will shell out for another monitor, or have one to spare. EMBM is currently only supported by the G400, and so game developers are unlikely to code for one card alone, regardless of what sort of visual quality is gained. If other cards supported EMBM (which would be nice) it probably would be implemented more, but then the reason for getting a G400 would evaporate.
Glaze3D
Bitboys' futuristic 3D chip offers incredibly fast 3D acceleration, with a 9 meg frame buffer on chip. However, it is at least a year away (I think) and so comes under the heading of vaporware, especially as Bitboys didn't deliver on their last product (I think). It may have a 1 GTexel or more fill-rate, except I can't remember. I don't think it features T&L either. As you might have guessed, I haven't been paying a whole heap of attention to the Glaze3D. Anyway, the 9 meg of onboard frame buffer will partly eliminate a bottle neck, that of the memory bus, which would be quite substantial at these fill-rates. Texture memory will still be off chip, allowing for flexibility for the manufacturer, as well as keeping the cost down (on chip memeory is expensive).
Conclusion
So what card should you buy? Well, the GeForce is available in Singapore right now, for about SG$529, although the drivers are still beta and it only has SDR ram. So it's not a future card, as the 
