What do you want to see the most in the next gen video cards

[rwolf]

What do you want to see the most in the next generation video cards?

 

[mat]

i voted for More Programmability.
I think its important to get a flexible enough System and then scale performance up (or do both at the same time) so Games dont have to be coded for 3 Architectures in mind.

i'm basicly thinking they shoud define a Standard like on CPUs and then only increase perfomance and add new stuff (like MMX/SSE/...) when needed (every 3 Years or so).

Edit: but since i dont do 3D-Apps, i have no idea if this is really a concern to coders

 

[Arun]

ILDP, TBDR, Stochaistic MSAA transforming aliasing in noise and ( 8 FP32 MADs OR 16 TEXs ) AND 8 COS/SIN ops/clock.

Okay, so we won't get all that next gen, but hey, we'll get it - evantually.

Although we'll get TBDR with Series 5, Stochaistic MSAA in NV40, and maybe the ops number in NV40 - although the NV40 pipeline organization remains a mystery...


Uttar

 

[keegdsb]

More innovation like Surround Gaming/Triplehead, (Matrox probably won't be back for another round until Longhorn) something that is unconventional, and unheard of.

Though on the "standard" front, not just faster AA/Aniso, but better quality.

 

[jvd]

I didn't vote. I want them all

 

[Ostsol]

More programmability, baby!

 

[Xmas]

PPP

 

[MuFu]

I don't care as long as the box is nice and colourful and has impressive numbers on it.

MuFu.

 

[tobbe]

Id like to run MAME on my next GPU

 

[Enbar]

I voted for more pipelines because it was the closest to what I want, but higher clocks would also help. What I really want is fast pixel shaders and I don't care how they get them. What I'd find useful is being able to cover the whole screen with at least 32 instruction shaders and then a small number of important objects with even better shaders. Once that's doable then we can work on more features.

ILDP, TBDR, Stochaistic MSAA transforming aliasing in noise and ( 8 FP32 MADs OR 16 TEXs ) AND 8 COS/SIN ops/clock.

Okay, so we won't get all that next gen, but hey, we'll get it - evantually.

Although we'll get TBDR with Series 5, Stochaistic MSAA in NV40, and maybe the ops number in NV40 - although the NV40 pipeline organization remains a mystery...


Uttar

I found uttar's thoughts interesting, but I don't agree. I have recently heard nvidia preaching the importance of cos/sin, but I suspect it's because it's one of the few instructions they beat ati at. I can remember them preaching other features in the past like paletized textures, which oddly enough their competition didn't support. Call me cynical but I've found nvidia building up questionable features just because they're better at them than the competition, and not becuase they are nearly as useful as they claim. Of course ati does it too, can we say trueform? In reality I find myself rarely needing a lot of trig functions and when I need one a texture lookup has always been good enough. Also I'll be disapointed if there are only 8 mads per clock.

 

[bloodbob]

Real time ray tracing
Real time radiosity / global illumination
Real time photon mapping and caustics

Thats what I want :/

Currently I wanna see more ansio and more AA and filrate because the progamablity is currently not saturated.

 

[3dilettante]

I want the full trilinear filtering present in cards from two years ago.
edit: as a given, not as something considered extra

As for the fuzzy future:
Maybe instead some kind of nifty programmable mipmap generation/replacement. Or a way for some of the geometry calculations to provide feedback to the CPU for collision detection, etc.

 

[vrecan]

I voted for more pipelines because it was the closest to what I want, but higher clocks would also help. What I really want is fast pixel shaders and I don't care how they get them. What I'd find useful is being able to cover the whole screen with at least 32 instruction shaders and then a small number of important objects with even better shaders. Once that's doable then we can work on more features.[

that is exactly what i want in next gen hardware, more shader speed! I didn't see anything that addressed what I wanted in this poll directly so I selected other.

 

[Ailuros]

Where's my "no more questionable per game optimisations" option?

Uttar,

Although we'll get TBDR with Series 5, Stochaistic MSAA in NV40, and maybe the ops number in NV40 - although the NV40 pipeline organization remains a mystery...

Just because NVIDIA includes TBDR as a possible option in it's PPT presentations to increase bandwidth, doesn't mean that they'll actually seriously considering it.

I'd like to remind you that the pipeline organization, as the utilized algorithms on both prementioned architectures are still a mystery. Pick on Simon for one of the candidates, but I doubt he'll be willing to tell anything to anybody ....

 

[Pete]

All of the above, please. And for less money.

 

[bloodbob]

Or a way for some of the geometry calculations to provide feedback to the CPU for collision detection, etc.

Yeah they keep talking about how you can use displacement mapping for terrain yet you still need to do all the tesselatins on stuff in the cpu because you need to do collosion detection.

 

[Colourless]

It would be interesting to see some basic physics test implemented, such as even something as simple as a per poly sweep test. How difficult something like that would be to implement in hardware, don't ask me.

 

[Ostsol]

Or a way for some of the geometry calculations to provide feedback to the CPU for collision detection, etc.

Yeah they keep talking about how you can use displacement mapping for terrain yet you still need to do all the tesselatins on stuff in the cpu because you need to do collosion detection.

All you need in system memory is the heightmap. Collision detection can be performed on the interpolated height at a given (x, y) location. Using floats it requires a third of the memory that a tesselated mesh would need and is probably faster.

 

[Tahir2]

All of the above, please. And for less money.

Preferably free. Brings a whole new meaning to the term 'free FSAA.' 8)

 

[KimB]

I voted for "merged pixel and vertex shaders," though I'm not sure that's the best wording. I think it would be nice if pixel and vertex shaders both used the same functional units, however. That should mean more efficient usage of available processing power, and, I believe, would be relatively easy to generalize to using the same hardware for primitive processing as well.

Edit:
Oh, btw, as a few pointed out (sort of), there was no mention of increasing AA/texture filtering quality, just performance. nVidia has a long way to go in AA quality, and I think ATI could stand to improve their aniso quality (while nVidia should go back to the quality available in the GF4).