The ffts take something like 20% of the time, but they are many short FFTs.
I am not sure if there are FFTs libs for ATI which can run with the openCL environment. Maybe Gipsel can tell us something.

gdf

https://developer.apple.com/mac/library/samplecode/OpenCL_FFT/index.html

here it is source code for a fairly complete library for FFT (made by Apple, but should be easily ported). Don't know about the license, could it be without any?

from what we have seen the performance on the HD4xxx series is very very slow.
So it is runs but very slow. The ATI openCL emulates the shared memory via global memory. At this point they are so slow that it is likely that we will restrict to HD5xxx cards.

gdf

Hm, if a HD48590 is really considered as VERY SLOW I maybe should completely forget about crunching.... :-(

---

Greetings from Cori

To quantify GDF's remark about the speed of 48xx cards, we currently find that a 4850 runs our code about as fast as an NVidia 8600. This is almost entirely down to the lack of local shared memory and we have every reason to expect that the 5xxx cards, which have shared memory on the processor itself, will be significantly faster. We're still waiting for the postman, though!

MJH

https://developer.apple.com/mac/library/samplecode/OpenCL_FFT/index.html

Thanks for the link, cenit. There are several FFT OpenCL projects we're watching - hopefully one will have all the features we need soon (In case you are wondering, these are: batched 1D r2c and c2r transforms with an FFTW-like interface and data-layout). Alas, we don't have the resources to develop one in-house at the moment.

MJH

from what we have seen the performance on the HD4xxx series is very very slow.
So it is runs but very slow. The ATI openCL emulates the shared memory via global memory. At this point they are so slow that it is likely that we will restrict to HD5xxx cards.

gdf

Hm, if a HD48590 is really considered as VERY SLOW I maybe should completely forget about crunching.... :-(

Or run out and buy 5xxx series cards ... :)

Then again it may be a "temporary" thing when ATI hears about it ... there may be a way to change the API so that this stops being an issue. Sometimes they can be real clever these software types. Remember the OpenCL is very new ...

Heck, we are a year plus into GPU computing on BOINC and it is still not working as well as it should ...

Is it slow because it uses system RAM or just does not have any GPU Cache RAM???

graphics dont really need cache so ATi can get away with out having it working for now. GPGPU is a very different story. btw, its not technically a cache. its called local data share. as of now only brook and IL have support for LDS. its the exact same reason folding@home is so slow on ATi cards. to give you an idea of how fast LDS is vram is 115GB/s and LDS is 1.9TB/s on a 4870, a lot of apps need that bandwidth which is almost 20x faster than main memory access.

Is it slow because it uses system RAM or just does not have any GPU Cache RAM???

the cache is there but it will take a driver update to get it working so for now its vram.

I updated beta drivers + OpenCL SDK as requested

Starting BOINC client version 6.10.19 for windows_x86_64
ATI GPU 0: ATI Radeon HD5700 series (Juniper) (CAL version 1.4.467, 1024MB, 1360 GFLOPS peak)
OS: Microsoft Windows 7: Ultimate x64 Edition, (06.01.7600.00)
Message from server: No work sent
Message from server: Your computer has no NVIDIA GPU

Exaclty the same here (except for having client 6.10.18). Same GPU/OS. Enabled doing Beta Work in the account settings. But still getting no work.
I'm looking forward to help you, but don't know how.

Johannes

I updated beta drivers + OpenCL SDK as requested

Starting BOINC client version 6.10.19 for windows_x86_64
ATI GPU 0: ATI Radeon HD5700 series (Juniper) (CAL version 1.4.467, 1024MB, 1360 GFLOPS peak)
OS: Microsoft Windows 7: Ultimate x64 Edition, (06.01.7600.00)
Message from server: No work sent
Message from server: Your computer has no NVIDIA GPU

Exaclty the same here (except for having client 6.10.18). Same GPU/OS. Enabled doing Beta Work in the account settings. But still getting no work.
I'm looking forward to help you, but don't know how.

Johannes

And even with my slow HD4850 I'd like to do some test WUs before giving up here but well... I don't get work either. *grin*
I've updated my account settings, drivers and the sdk package but no success.

---

Greetings from Cori

We are not giving work yet. Waiting for a ATI5870 to test.
gdf

Any updates? Did you contact PoorBoy on his generous offer? This is the project I wish to see 5xxx implemented, then I will go and buy these cards. I'm not prepared to do that just for mathematical problems (Collatz), but this project and the science made makes it almost sound smart ;)

We could not find a 5870, but we got a 5850 that should be good enough.
I have just installed now, removing the old 8800gt.

gdf

Would any of the WU's require double precision? The new 57xx's apparently don't have support for it...

We could not find a 5870, but we got a 5850 that should be good enough.
I have just installed now, removing the old 8800gt.

gdf

I find this interesting that in the USA there is no ATI 5800/5900 GPUs
Even in far distant New Zealand I can get ATI 5970 in 3 days.
I would love to crunch GPUGRID again but had to give up with a 295.
I am cruching Collatz now because a 5790 can rip though the WUs 2 every 7 mins.
So come get a optiz - app for ATI cards . You are missing out on so much Computer Power
Ross

---

We are working on the optimization now. We will not put the application out until it delivers what a 5870 should deliver in terms of performance. Be patient.

gdf

We are working on the optimization now. We will not put the application out until it delivers what a 5870 should deliver in terms of performance. Be patient.

gdf

Some of us are getting too old to have Patience ... :)

We are working on the optimization now. We will not put the application out until it delivers what a 5870 should deliver in terms of performance. Be patient.

gdf

Great, I will be patient!
Ross

---

Well, if the performance of 5850(70) would be equal to the performance of nVidia top devices, this optimization will be successful, otherwise no one would use their top ATI cards with PPD of 2 years old nVidia. But honestly I hope ATI would outperform nVidia in this very important project. Good Luck Fellows.

Despite credits GPU-Grid has one major advantage compared to both other BOINC projects for ATI: it's got a reliable server and enough work. If MW and Collatz were running fine, they'd offer enough work as well.. but for now MW still crashes rather often and when ever this happens the Collatz server is brought to its knees. GPU-Grid could bring some balance to the force ;)

MrS

---

Scanning for our furry friends since Jan 2002