Any inside(r) infos about crunching with ATI cards ?

[FalconFly]

...I'm curious because I've re`d e upcong BOINC Versions supposedly support OpenCL (work in progress).

Does anyone have some clues as to how far (if at all) there's a cooperation with ATI on this?

[sunu]

As far as I know, ATI hasn't even released an opencl driver for developers yet. On [wsl=h4tp://f��ums.amd.com/devforum/messageview.cfm?catid=328&threadid=112397]this[/url] forum post, AMD employees say their opencl implementation will be publicly released in the second half of 2009.

On the other hand nvidia has already releed `n opencl driver to registered developers ( http://www.nvidia.com/object/cuda_opencl.html ).

[arkayn]

So far the only project that can crunch with ATI cards is MilkyWay, and that is via the anonymous platform option.

And work is scarce at MW at He moment.

[FalconFly]

Ro I guess it'll be some time later this year until we see ATI cards supported by a certified OpenCL architecture under BOINC...

Might become an interesting year after all, GPU crunching finally catching up

[chelski]

So I guess it'll be some time later this year until we see ATI cards supported by a certified OpenCL architecture under BOINC...
Might become an interesting year after all, GPU crunching finally catching up

From the feedback of those who runs folding@home, ATI cards tend to underperform similar nVidia cards at folding even though the theoretical FLOPS are higher.  There's plenty of work to be done before ATI is even competitive to CUDA which seems to be more widely supported, but I suppose their focus is on the graphics rather than the GPGPU side.

[arkayn]

I have a feeling that once Milkyway gets the GPU side running, ATI will blow Nvidia out of the water.

[popandbob]

ATI has gone to single precision for CUDA which will be a real advantage. In single precision mode both nvidia and ati are theoretically equal. This also means that once the ATI is switched to single precision we will see some crazy number crunching as both cards offer much higher flops in single precision.

[Gecko_R7]

.... This also means that once the ATI is switched to single precision we will see some crazy number crunching as both cards offer much higher flops in single precision.

Yes, I'm also waiting to see what Gipsel/Cluster Physik is able to make from Travis's code revision.

However, remember that Travis's reported times for his Cuda app was on a 9600M that has only 32 shaders & relatively slow 1075-1200Mhz speed.
Would be interesting to see an apples to apples w/ @ 4850 vs GTX 285 for example.  Both cards are @ 1 Tflop performance.

[arkayn]

My 4830 does pretty good on MW, when it can get work.

[_heinz]

which driver do they use in the milkyway ap now?
have latest Catalyst 9.4 for Vista64 from ATI, did not work on milkyway
did they alredy solved  the problem to run BOINC as a service ?

[arkayn]

I ran both 9.4 and now 9.5 with my 4830.

The big thing is the renaming of .dll files that AMD did, we had to do this to get it to work.

Copy

aticalrt.dll or aticalrt64.dll
aticalcl.dll or aticalcl64.dll
aticaldd.dll or aticaldd64.dll

Now Rename the copies to

amdcalrt.dll or amdcalrt64.dll
amdcalcl.dll or amdcalcl64.dll
amdcaldd.dll or amdcaldd64.dll

Just like CUDA, you cannot install it as a service.

[_heinz]

Hi arkayn,
to which directory I must copy and rename the DLL's
BOINC ?
ahh.. C:\Windows\System32

[arkayn]

Yep, forgot to put that one in there.