ghost_zero5

OK. I have now done further testing.
I tried the 196.21 with the mod and though GPU-Z reports "no CUDA" I can perfectly run programs like:
CudaMFLOPS1 or Cuda Memtest and CUDA-Z. All of them report one device with CUDA capabilities and use that one.

Furthermore, I tested the 185.68 driver and there GPU-Z also reports "no CUDA" but as with 196.21 and the mod I can run those three programs just fine.

Which programs did you test?

Kenpojujitsu3

I saw that neither I nor that youtube user had AA enabled per the benchmark settings in Batman (though I have no idea what he did in CCC or behind the scenes) so I tested some OC'ing last night into this morning. OC'd from stock 2.8Ghz to 3.8Ghz (LinX stable but fails Prime95 after 2 hours so still working on stability). Got some new and somewhat interesting numbers with the approx. 35% clock increase.

Batman Arkham Asylum Benchmark

Crossfire + Physx performance went up to 98fps from 77fps (27% increase). Min went from 41 to 58fps (41% increase).

Fluidmark Benchmark

Crossfire + Physx Performance went up to 102fps from 66fps (54% increase).

Star Tales Benchmark

Crossfire + Physx Performance went up from 19fps to 57fps (200% increase).

The CPU defintely is having something to do with the Crossfire + Physx performance. Looking at the Star Tales Benchmark tripling it's framerate, I'm inclined to say the system was bottlenecked at 2.8Ghz (or Intel Speedstep which I had on at stock settings but off for OC'ing was not triggering the full clock rate of 2.8Ghz). I'm going to do some more testing when I get out of the office and over the weekend (like running at 2.8 with Speedstep OFF). It may prove useful to know just how much horsepower is needed on the CPU side to use this kind of setup effectively.

GenL

Kenpojujitsu3, thanks for a very interesting info.

CPU overclock should really help with fps even for Crossfire for some CPU-hungry games, but in my opinion CPU clock increase should do nothing to Crossfire when you are using PhysX, because you have 5870, which is more than enough for games like Batman - because the limiting factor is PhysX performance, not a 5870 Crossfire.

So, maybe it's all about PhysX, which uses much CPU power even in GPU mode? I mean, hardware PhysX was always meant to use PPU/GPU to take off the load from CPU, and if it's true we shouldn't see such increase after CPU overclock. Looks really interesting (to me at least).
More tests and comparisons with CPU overclock + PhysX on/off states would really help here. I hope to see more from you, Kenpojujitsu3, thanks.

__________________
I don't want to support PhysX, but i'd like to support people who want to utilize their hardware the way they need/like.

Kenpojujitsu3

Either my setup is weird or Nvidia is selling a bunch of nonsense when advertising just how much Physx is offloading to the GPU. It may be offloading but it is still VERY CPU dependent by what I'm seeing.

I dropped my OC to 3.59Ghz where it's LinX and Prime95 stable and ran some more benchmarks.

Batman Arkham Asylum Benchmark

2.8Ghz - runs at 77fps in single and crossfire mode.
3.8Ghz - ran at 95fps, only tested crossfire mode.
3.59Ghz - single card 83fps, crossfire 91fps.

Looking at how a higher clock brought separation to the single card and crossfire modes here I'm gathering that it's possible physx is taking up a considerable amount of CPU cycles which was causing a bottleneck at 2.8.

Fluidmark Benchmark

2.8Ghz - 66fps, single card & crossfire
3.8Ghz - 102fps, crossfire test only
3.59Ghz - 112fps, single card and crossfire

I'm going to chalk this up to that 3.8 OC being known to be unstable. But still, noticeable increase with CPU clock and I'm positive a 5870 or 2 is not the limiting factor when rendering this.

Star Tales Benchmark

2.8Ghz - Speedstep and Halt State enabled, Crossfire 19fps
2.8Ghz - Speedstep and Halt State enabled, single card 88fps
2.8Ghz - Speedstep and Halt State disabled, 27fps, crossfire test only
3.8Ghz - 57fps, crossfire test only
3.59Ghz - single card mode, 112fps
3.59Ghz - crossfire mode, 40fps

Still interesting how this one behaves between single card and crossfire. Again, performance scales up with CPU clock speed and is impaired when instructions need to be sent to another rendering card.

Seems to me Physx isn't as CPU independent as Nvidia would have people believe. All the more reason for open physics implementation.

Edit: Batman runs at 155fps on a single card and 266fps crossfired when Physx is off. From 155/266 to 83/91 is a hell of a performance drop with a dedicated Physx card considering that Physx is "supposed" to free system resources and rely on the GPU and Video memory of the dedicated card don't you think?

CaptainFabulous

I tried those programs you listed and they do seem to work. CUDA-Z reports it's there. CudaMFLOPS1 seems to work as does an old version of CUDAMemTest.

Matmul spits out this:

There is 1 device supporting CUDA
GPU 1: "GeForce GT 220" SM:1.2
MP: 6 Cores: 48 Freq:1.57 GHz
Mem: 1024Mbytes BW (Pinned) H2D: 2.40GB/s D2H:2.79GB/s D2D:18.17GB/s

CUDA floating code
Time running: 1.626000 s. Gflops: 84.515484 Gflops.
CUDA no IO floating code
Time running: 1.510000 s. Gflops: 91.008064 Gflops.
CUDA double code
!!!! kernel execution error.
CUDA no IO floating code
!!!! kernel execution error.

And Badaboom errors out with this:

GPU 0: ATI Radeon HD 4770
GPU 1: NVIDIA GeForce GT 220
FATAL:Could not get compute capability from device driver
WARNING:HW Decode Capability detection failed.
FATAL:Could not acquire CUDA device.
FATAL:

Also, CoreAVC won't run in CUDA mode either. BOINC reports there is a CUDA device but I'm not entirely sure it's being used even though I've told SETI to use it.

EDIT: after doing some updating and altering preferences SETI is indeed running on CUDA. So I guess it's possible, but some programs have problems with it.

ironleg1963

after reading to page 78 i found some files to fix my agiea card. is there updated files in one place. ( 108 pages is alot to read) I have an agiea physix card bought directly from them and a asus ati 4870 x2. I am willing to test drivers to help fix this thing. nvidia is not thinking clearly on this- its like buying a car and not driving it because the engine is proprietary and they wont let you have fuel to run it.

ghost_zero5

Now the question would be why. Is it just because it isn't the primary card that those programs have troubles with it or are the CUDA capabilities limited when it isn't used as primary card (or because of the ATi card).

R3nzoku

I have read all 108 pages, tried all combinations, and I cannot get nv cp to show me the physx config.

Should I reinstall windows? Any workaround?

PLEASEEEEE HELP ME!

Attached Thumbnails

 

Alexandr82

Minimum of 32 cores and a minimum of 256MB dedicated graphics memory for PhysX!

GeForce 8400GS - 16 cores

R3nzoku

yes but why does it work when I have a 8600gt as a primary card?