Wingz-
We are all over in the Video forum, where have you been???
Soccerman-
"I've seen a few (in picture format mainly), and they look quite nice, however their polygon models aren't much more complex then, say, Quake 3."
Check out Giants. While the polygon count can be had on a non T&L board, the FPS suffer quite a bit, particularly with "slower" CPUs(sub GHZ are quite slow).
"I don't think they're running that many more"
Most of them aren't. We only need one single title to show that the unit works as advertised, and there are plenty more then that. With scaleable LOD becoming increasingly popular, it will be much harder to get accurate comparisons without decent benchmarks, but there are developers on record stating that they run both higher FPS and higher polygon counts then software T&L on today's CPUs. The detractors said repeatedly that the GeForce's T&L unit couldn't compete with then current CPUs(November '99) which is laughably inaccurate.
"It runs on Direct X 8 and it's pretty nice, but the order of magnatude of polygons compared to the 'previous generation' of games (like Quake 3) isn't anything more then double, more likely 1.5x. I wonder what kind of speed is guaranteed from a Geforce DDR?"
That game isn't a good example(do you think that looks good? Visuals look quite poor to me compared to some games already availalbe). When the game based on Crytek's engine comes out it will likely be a good showcase. Also, you can also look to the NV15 level when paired with a PIV and dual channel RAMBUS. Without even using the hardware lighting engine it is quite playable(60FPS) and is cloer to a five to six hundred percent increase over a standard Quake3 level.
"btw, these new games aren't entirely T&L dependant, right? You'd still see a difference between, say a 400mhz and 600mhz CPU right?"
You will see a bigger difference then without T&L. Offloading the T&L strain from the CPU the limiting factor becomes the game code for anything out now or in the not too distant future. Because you reduce the amount of strain on the CPU, the 600MHZ CPU will have a bigger framerate increase when using hardware T&L then not, until you become either memory bandwith(system or local) or T&L limited(no game is close to being T&L limited yet).
Figure a game devotes 30% of it's time to T&L when running in software mode and runs at 50FPS on a 400MHZ CPU. If you calculate for a 600MHZ CPU then you will be at roughly 75FPS, moving T&L off of the CPU and on to a dedicated unit will drop this down to closer to 5%. So, the 400MHZ CPU would be hitting roughly 62.5FPS while the 600 would be hitting about 93.75FPS.
We would need an explosion in poly counts to change the type of scenario above. A faster CPU should offer you even greater benefits with hardware T&L then without.
"If I could get a Kyro 2 to run that, I would! c'mon PowerVR, I'm counting on you!"
Are you talking about the game you linked to? Well, from looking at the screenshots the visuals are quite drab. Lousy texturing, the models aren't very well done although there are some nice particle effects/explosions. If you are thinking that a deferred will help you much, I wouldn't hold your breath. From looking at the screenshots it looks like this game will have maybe 2X, if that, overdraw the overwhelming majority of the time. Without a sizeable amount of overdraw the KyroII is likely to get manhandled by pretty much any other board that is current when it launches. This may not be the case, but I'm hearing 166MHZ core clock, unless there are some serious changes from the Kyro1 then it simply won't be competitive particularly not in a game like the one you linked to.