We are practically ready with the app alone. We have to test it on the context of boinc now.

First thing to do is to update the server to a more recent version, then wait for problems. Our server software is quite old and there could be problems which would only appear in a couple of weeks. I don't want to test AMD with the old server as no opencl was there then.

In the meanwhile we will do our tests locally on a private new server.

An other thing to say that it will be windows only for the first several months.

We have tested on one OS (XP32) and one card (H5850). So expect few issues. Hopefully, it all works and we can start submitting real workunits for amd. In such case they will be on a different app, not to affect the work on nvidia cards, but the porting of the code is complete, so they will be of the same usefulness of nvidia wus.

gdf

The new amd app will require OpenCL 1.1 and double precision cards?

Is it possible to update the server for OpenCL without the *new credit* code?
If not, I would like to suggest that you run some NVidia WUs against the new server and let us know what we might see for credit per WU.

Not that I am participating in this excellent project for the points but we all know how fussy crunchers get when it comes to credits :-)

---

Thanks - Steve

The new amd app will require OpenCL 1.1 and double precision cards?

The nVidia code doesn't require dp, so there's no reason why the ATI code would. Can't tell you the OpenCL version, though.

MrS

---

Scanning for our furry friends since Jan 2002

opencl 1.1 and jut single precision required.
H5000, H6000 and the new H7000 cards should work.

gdf

opencl 1.1 and jut single precision required.
H5000, H6000 and the new H7000 cards should work.

gdf

...it's time to burn some fusion chips
From my point of view, it doesn't matter if the gpu load is going to be 20-25%, I think my laptops are going to handle it

gpu load is around 85%

g

gpu load is around 85%

g

any rough estimate for alpha testing?...like next week? :)

...it's time to burn some fusion chips

The current Fursions are HD6000 series, so they should work :)
(don't expect performance miracles, though)

MrS

---

Scanning for our furry friends since Jan 2002

gpu load is around 85%

g

any rough estimate for alpha testing?...like next week? :)

More like next month.

gdf

gpu load is around 85%

g

any rough estimate for alpha testing?...like next week? :)

More like next month.

gdf

Ok. Now is 1.10.2011. What you waiting for?

What you waiting for?

Maybe for the app being ready for public alpha testing?

MrS

---

Scanning for our furry friends since Jan 2002

We are waiting to have time... We are submitting 2 new papers and two new applications for grants.

gdf

Any news on the AMD application?

I have two AMD GPUs I would love to use on this project! I hope you can get it out soon!

I am also interested by the future support for ATI / AMD, but I hope it will be stable and the performance will be a good level ?

I am part of the beta-team on folding @ home and for this project the performance of ATI / AMD is not very good even if there are improvements.

It's hard to compare performances between different projects, especially Boinc to non-Boinc. However I would still expect a high end AMD GPU (HD6900) to significantly outperform a high end CPU (i7-2600), which is not the case at folding.

We know the performance is not great compared to high end GTX500 series GPU's by NVidia, but again we are comparing mature CUDA apps to a pre-Beta OpenCL/GL app.

Basically a GPUGrid entry level NVidia card (say GT240 to GT430) would be on par with a mid-range AMD GPU (say HD5850 to HD5870). Obviously the high end 6000series should fair much better; perhaps approaching say a GTX470. I would expect the high end Southern Islands GPU's (7000 series), supposedly due out before the end of the year to at least match a GTX470. That said, Kepler would stretch the gap again, when released.

Would be beta testing will come down to time/timing; some work presently under way has to come first. How long that takes depends on the problems that are encountered. Not sure if a new NVidia app will happen first (presently on hold until next 2012), perhaps just for Linux, what the situation is WRT publications, seminars..., and the researchers might want to finish off some work using older code before anything happens.

Should a beta test run, be successful, and a live app be successfully used for a while, development might mean the gap here between AMD and NVidia apps would fall for GPUGrid, perhaps even with a mature OpenGL 4.2. However, such positive speculation should be made cautiously; a Beta could demonstrate unreliability, and require too much work from the researchers to get an AMD project off the ground.

---

FAQ's

HOW TO:
- Opt out of Beta Tests
- Ask for Help

AMDs current VLIW architecture is much harder to utilize efficiently than nVidias scalar shaders. That's likely why they've got these problems at F@H and why they're not yet rocking GPU-Grid. Sure, software is also a very important piece of the puzzle.. and AMDs compiler has a harder job keeping these execution units fed. But it's not always possible to work around fundamently architecture limitation by software - so it's not only about software maturity and "trying hard enough".

AMDs upcoming "Graphics Core Next" (GCN) will be the first departure from the VLIW since many years. It should be a lot easier to utilize. However, I doubt the same code path will run good on both, GCN and VLIW. That will lead to a very interesting, if not at times frustrating situation.

BTW: actually there's not that much of a difference between Cypress (HD58x0) and Cayman (HD69x0) apart from switching from VLIW5 to VLIW4. At best Cayman is 20% faster per clock. That matters a lot if you're competitive, but is nothing if you've got a catch up an entire order of magnitude :p

MrS

---

Scanning for our furry friends since Jan 2002

Yes I think that the performance gap between AMD and Nvidia should be large enough.

when is that AMD GPUs will be supported ?

Cayman may only be 20% faster than a 5850 per clock, but there are some monsters compared to a humble 5850. For example, you beast cuts through MW tasks in around a minute, while an HD5850 takes about three times as long. I'm not saying AMD's will be great, but some people might want to use their AMD GPU here.

If an HD 5850 really is about 1/4 the speed of a GTX580, then your Cayman might be somewhere around 2/3rds the speed of a GTX580.

Even if performance is poor compared to a GTX580, there are other considerations:
People with existing AMD GPU's have expressed an interest in using them here.
Their performance per Watt or per Euro/$/£ is not known for the app.
Some ranges of cards might be more competitive than others; mid-range AMD GPU's would be competing against many CC2.1 cards, rather than CC2.0 GPU's.

So, worth testing sometime I think.

---

FAQ's

HOW TO:
- Opt out of Beta Tests
- Ask for Help