Future Hardware

[rwolf]

Any new features or ideas for next generation hardware beyond R600 and G80?

 

[BRiT]

They will be on red and green PCBs respectively.

 

[icecold1983]

id like to see af higher than 16x, and have usable 4x+ ssaa.

 

[Andrew Lauritzen]

Programmable interpolators, or alternative a negligible performance hit for doing it in the shader.

 

[Kanyamagufa]

Air conditioning.

 

[rwolf]

id like to see af higher than 16x, and have usable 4x+ ssaa.

Any idea when you reach the point of diminishing returns?

 

[nAo]

Programmable interpolators, or alternative a negligible performance hit for doing it in the shader.

log shadow maps here we come!

 

[icecold1983]

Any idea when you reach the point of diminishing returns?

id say 128x af and 16x ssaa would be good for the foreseeable future.

 

[Andrew Lauritzen]

log shadow maps here we come!

Oh yeah, hadn't even thought of rasterizer interpolation... that would certainly be powerful if it was programmable. I'm not sure how efficient that could be made, but hey, this is a dreaming thread right?

 

[Jawed]

You could do your own arbitrary rasterisation in the GS, couldn't you? Convert an input triangle into "single pixel" triangles - spacing them however you want, resulting in your log-space interpolated rasterisation.

Apart from being slow you'd also have the problem of limited amplification within GS, so I guess you'd have to multi-pass input triangles until each one is entirely GS-amplifiable.

Jawed

 

[fellix]

More streamlined caching system, flexible virtual memory addressing, some push on eDRAM development and implementation (stacked cells or sort of).

 

[nAo]

You could do your own arbitrary rasterisation in the GS, couldn't you? Convert an input triangle into "single pixel" triangles - spacing them however you want, resulting in your log-space interpolated rasterisation.

Apart from being slow you'd also have the problem of limited amplification within GS, so I guess you'd have to multi-pass input triangles until each one is entirely GS-amplifiable.

Jawed

I would not call it (super)slow, I'd call it impratical..or maybe even impossible. Can you have a single GS invocation generating milions of primitives?

 

[Jawed]

GS is limited to outputting 1024 32-bit values per input primitive, so it's definitely not many output vertices :smile:

Jawed

 

[trinibwoy]

id say 128x af and 16x ssaa would be good for the foreseeable future.

Given that the difference between 8x and 16x AF is already miniscule I think we're pretty much nearly maxed out there.

 

[zgemboandislic]

Triple slot, 500w, 15" long

 

[nonamer]

DX10.1: Geometry shaders that don't suck?

 

[icecold1983]

Given that the difference between 8x and 16x AF is already miniscule I think we're pretty much nearly maxed out there.

in racing games, 16x isnt nearly enough in many scenes

 

[nAo]

in racing games, 16x isnt nearly enough in many scenes

That's developers fault, using AF to make roadsigns/marks not blurry is like using a nuke to kill some flies

 

[rwolf]

DX10.1: Geometry shaders that don't suck?

Why do geometry shaders suck?

 

[Shtal]

64bit color (Billions of colors) vs. we have 32bit (16 millions), I know Linux has in their software 64bit, but I like to see in hardware.

Something like this "resolution" 2560x1600x64bit