1.
Interesting review. Could you also include FX5900 results for comparison to R9800Pro? FX5900 would also have shown how a 4x2/8x0 pipeline compares. It would also be interesting to see what a less memory bandwidth R9600XT could do.
Unfortunately, Nvidia's shading/scheduling driver optimizer is treated with too much cynicism. The FX architecture has real VLIW programming issues that although ps_2.0->ps_2.0_a HLSL profiling helps, the run-time optimizer is still required. Of course we assume coarse cheats are now gone...;)
The real issue of 52.xx drivers wasn't explored. The reason filtering performance is reduced to pseudo-trilinear for all D3D & shader performance is improved is because the FX architecture shares the texture engine for pixel operations. So reducing tex load (regs) allows better fragment pipeline performance. Also by reducing trilinear filtering load to single-cycle bilinear, you don't pay the penalty of 2x2 multi-texture operation & the GPU can operate in 4x1 mode more of the time. ATI's architecture can simultaneously/independently tex/ps to is unaffected - although R9600 did/does(?) have a pseudo-trilinear filtering mode, too.
Unfortunately, Nvidia's shading/scheduling driver optimizer is treated with too much cynicism. The FX architecture has real VLIW programming issues that although ps_2.0->ps_2.0_a HLSL profiling helps, the run-time optimizer is still required. Of course we assume coarse cheats are now gone...;)
The real issue of 52.xx drivers wasn't explored. The reason filtering performance is reduced to pseudo-trilinear for all D3D & shader performance is improved is because the FX architecture shares the texture engine for pixel operations. So reducing tex load (regs) allows better fragment pipeline performance. Also by reducing trilinear filtering load to single-cycle bilinear, you don't pay the penalty of 2x2 multi-texture operation & the GPU can operate in 4x1 mode more of the time. ATI's architecture can simultaneously/independently tex/ps to is unaffected - although R9600 did/does(?) have a pseudo-trilinear filtering mode, too.
[Posted by: kazzaa
| Date: 11/04/03 08:33:34 AM]
| Date: 11/04/03 08:33:34 AM]


