Article on CineFX/NV30 Architecture (English)

[Demirug]

The big question on my mind is how much performance is NV30 letting go un-utilised because its drivers just aren't sophisticated and clever enough?

Will the Detonators 50.xx raise performance by 2% or 20%?

Is the design so complex that only Stephen Hawking's could solve how to optmise NV30's drivers?

What do folk hope/expect/wish we might see realised with the Det 50.xx series drivers?

Playing a little bit around with the new shadercompiler from the DX9 SDK I see that some shaders runs up to 40% faster if you use the new PS 2_A profile. But there are other shaders that run at the same speed as before.

 

[991060]

I dont understand why you have to pack the tex and arith inst to favor R300/350.

won't the arith unit be idel when the shader core is performing the tex operation?

 

[Demirug]

I dont understand why you have to pack the tex and arith inst to favor R300/350.

won't the arith unit be idel when the shader core is performing the tex operation?

No, AFAIK the two units working on two different pixel at the same time.

 

[991060]

No, AFAIK the two units working on two different pixel at the same time.

if so,should the sequence of the shader be important?

can the R300/350 shader unit do tex1 op on pixel1 and pixel2 or tex1 op and arith1 op on pixel1?

 

[PSarge]

I dont understand why you have to pack the tex and arith inst to favor R300/350.

won't the arith unit be idel when the shader core is performing the tex operation?

It might be that you want to issue all of your texture fetches as early as you can. If you do that then you can be doing other things until you actually use the texture result, at which point you'll stall until the texture fetch completes.

This is fine if the achitechture is happy having a large number of texture reads outstanding. If not then having a block of all the texture reads together at the start would mean the limit could be hit quite quickly. Then each texture fetch would cause a stall as it waited for a previous fetch to complete.

Pure guess.

 

[KimB]

I think the problem for nv3x was that nVidia did not start with a clean sheet...Had they had some foresight they'd have quietly scrapped nV3x last August and started over at that time.

Um, no. That would have meant that nVidia would have had nothing more than a GeForce4 (or a tweaked GF4) until the NV40. That was the trap that 3dfx fell into. nVidia was smart to get the FX out in some form as quickly as possible.