R300 ~20% Faster Than Ti4600

[blawson]

Haven't heard much about the new card, but I understand something to start looking for will be programmable vertex and pixel shaders, so the games companies can optimize the card for each game. Anyone know if this is so for these cards?

 

[jbond04]

Originally posted by: Soccerman


first of all, I've only actually done research on the high end 3D cards such as the Wildcat 3, and FireGL, which seem to me (at least now, cause it's obvious I've missed some) to have been the cards that are used for making high polygon models. obviously there must be a level of cards used for even MORE intense situations if there actually exists a card that does not only T&L but Raytracing in hardware (whether it's FPU based or not I don't know).

how hard IS it to design a circuit specifically to do Raytracing? if they've had the ability to do it for a while (at least, at the speed necessary to do it in a game and make it look somewhat better than the best shadows), why haven't they done it already! is it just the speed that is still the problem?

I expect Raytracing to appear on boards at the level of the Wildcat 3 before I see them in consumer cards btw.

Cards like the Wildcat, FireGL, and Quadro are meant to assist in the display/rendering of the viewports in the 3D application, not the actual rendering of the scene (which is done by the CPU). Those professional-level cards do not help rendering at all.

However, a 3D accelerator for games is the renderer, so it is more likely to receive the raytracing enhancements, because it would probably be a good while before raytracing would be used in a viewport. (Hell, they are just finally starting to use Direct3D/OpenGL in viewports to the point where it has a benefit)

The reason that raytracing hasn't come down to the consumer level is because the processors are so expensive (and they're not fast enough). If you look at the process that the "Pure" (the professional hardware raytracing card) was manufactured on, it's .22micron! The last time a graphics card used that process was in 2000! Moreover, the Pure uses several of these processors in parallel (raytracing is greatly accelerated with parallel computations), so it is a very complex board to make. Plus, it uses proprietary plug-ins to interface with the 3D animation software.

If nVIDIA leveraged the power of a graphics card like the nv30 (which is just a big parallel processor, like all graphics processors) with their "Cg" programming language, I'm certain that raytracing calcs coud be greatly accelerated. The downside of this, of course, is that only a VERY powerful card could perform such calculations, so it's unlikely that we'll see any actual games making use of it for quite a while.

 

[Soccerman]

hmm lots to respond to..

first of all, I say rumored, because I don't trust websites that give out info before the NDA is up. it's a simple yet effective rule that I follow to prevent myself from getting caught up in specs of non-released cards.

so I call it (you're referring to the 20% number) a rumor.

now as for the performance increase argument, I'm talking about todays games, that includes SS:SE, Quake 3/RTCW, Mirrowind and so on.

You are far underestimating just how much faster the next generation of cards will be.

are you saying that the mere increase in transistor count is what is making you that confident that you'll say that?

their overall power will be vastly increased, but that is simply becuase they need it for the new DX9 functions. don't think this extra power is all going towards more pixel pipes, or floating point vertex/pixel shaders. what do I expect out of the NV30? a full DX9 chip, which means it'll be similar to the P10 in that the vertex/pixel shaders are FPU units, requiring alot more room to accomplish the same thing (though adding versatility also).

yes I know the P10 already doubles the vertex shader speed in theory of the GF4 by including enough FPU units to double the throughput, or even more because of finer granularity (ie, one instruction per FPU unit per clock rather than a fixed unit that can do 4 instructions per clock which will be inefficient when the program only asks for 3 or less instructions to be executed.

I expect all of these cards (P10, R300, NV30) to have at least double the shading speeds. that means alot of FPU units, which means alot of space being taken up.

and if ATi and nVidia want more efficiency out of the memory, they'll have to further advance their rendering designs, potentially eating up even more die space (perhaps the NV30 will also have a Command processor).

100% faster in RTCW at 1600 eh? that's interesting. I wonder if that's simply due to more efficient memory management/usage.

the R300 is expected to have 256 bit VRAM and it'll be faster than the Ti4600's memory clock,

that may be true, but I still think it won't be so tangible in DX7 type games like RTCW (yes I know RTCW is OpenGL, but I don't think they have any DX8 level features in it now do they?).

maybe I'm just underestimating what that extra memory bandwidth will do (maybe it's a necessity for upcoming games, what with so many passes and all).

Not necessarily and in fact it can be quite the opposite - complex pixel/vertex operations that execute multiple times on single pixels/vertices have the effect of shifting the bottleneck entirely onto the core and leaving the memory bandwidth alone.

actually they (the pixel/vertex shaders) have a limit to the amount of operations they can execute (I think it was 256 for the P10). while that is high by todays standards, there is a limit.

 

[merlocka]

Why is that dumb? My Ti4600 is more than 100% faster than my Ti500 in certain games at 1600 x 1200 (RTCW for example) and it doesn't have anywhere near the improvements for speed that a R300 will have over a Ti4600.

Is that with AF or AA on? because based on published benchmarks it looks like in RTCW 1600x1200 that the Geforce4 4600 is 50% faster than the Geforce3 Ti 500 which is about 50% faster than the Geforce2 Ti. This data is from the Digit-life 3Digest.

Of course, the Radeon 8500 (R200) is 300% faster than the Radeon 64 (R100) and about 90% faster than the Radeon 7500 in this same benchmark. It seemes to me that ATI's jumps have been quite large (Rage128 -> R100 -> R200) because they want their parts to last longer, as opposed to nVidia who refresh every 6 months even if it's just a clockspeed increase.