View Full Version : Tressfx needs CUDA optimization, SLI and tesselation also causes issues

5th Mar 2013, 05:42

SLI EVGA GEFORCE GTX680s (2x), running most recent WHQL drivers (will try new betas from today shortly)
win 8 pro 64bit,
32gb DDR3 corsair,
i7 3770k oc'd to 4.5ghz with corsair h100
Win 8 on its own Samsung ssd, tomb raider on 1 of 2 terabyte sata 3 drives
intel dz77ga-70k extreme SLI board with latest BIOS
running at 2560x1600, everything maxed...

experience extreme TressFX hair fps drop like many other NVidia users are already reporting.

I know that it's the TRESSFX hair that's doing it... as soon as I switch back to normal hair mode, FPS sky rockets. I know that TRESSFX uses more direct-compute tech than CUDA. No reason this should be choppy during upclose hair scenes on SLI'd 680 or 690 systems with everything maxed.

Please add CUDA support for TressFX.

Also, many are reporting SLI gtx600 series crashes/lockups with tessellation on.

5th Mar 2013, 05:48
I'd like to chime in with the nvidia tessellation with SLI crash issues.

But also, TressFX does not function correctly with SLI enabled. Hair behaves jittery and floaty compared to single GPU.

5th Mar 2013, 10:43
The up-close thing is an issue with many procedural effects done on video cards. I have no doubt they can fix it up for TressFX, but it's certainly not an Nvidia-specific issue.

Rewriting it for Cuda might help a bit, but the problem in your case is largely the architecture used by the GTX 600 series. TressFX requires a lot of general-purpose compute power, which Nvidia de-emphasized with the 600 series (as opposed to a large emphasis on it in the 500 series).

Also, I wouldn't expect to see AMD to rewrite it for cuda. Cuda was made by nvidia, and only their hardware supports it. Directcompute is supported by all DX11 hardware.

5th Mar 2013, 12:35
Rewrite the effect to CUDA will not help. DirectCompute is a fine choose for TressFX, because NVIDIA support this API correctly. Altough TressFX has some rendering problems on Geforce, but it will be fixed in a future driver.
For the performance ... well it's more like a hardware problem. I know this sounds silly with two GTX 680, but the NV Kepler architecture is not designed to effectively handle these complex workloads. AMD use a much better architecture in this perspective.

5th Mar 2013, 12:43
Guys, TressFX is made buy AMD, so no they are not going to optimise the drivers and hardware of nVidia hardware, it uses openCL to do it's work, and the reason your running into issues is that CUDA does not equal DirectCompute, infact nvida kinda destroyed the dc performance on most of thier GPU (them wanting everything to be physix and all)