I've just installed this version on an iMac running OS X Lion (10.7.2). The host is a late 2009 27" iMac with an E7600 (3.06 GHz) Core 2 Duo CPU and an ATI HD 4670 GPU. When using "Apple Logo -> About this Mac ...", the GPU is reported to have 256MB RAM. When BOINC is started it reports:

OpenCL: ATI GPU 0: Radeon HD 4670 (driver version 1.0, device version OpenCL 1.0, 512MB, 512MB available).

Has any body got any information about why the discrepancy in RAM size and if it matters in any way - ie, is it 'cosmetic' only?

I was originally intending to use this host for testing the new OpenCL app for Einstein. The Albert@Home project has been set up to do the testing and after doing a bit of reading there it seems that 512MB may not even be sufficient anyway. So I decided to give the OpenCL app at Collatz a spin. After attaching, files and a task are downloaded but when the OpenCL app actually launches, the whole machine locks up. I'm wondering if the cause of this might be related to the apparent discrepancy in RAM size.

Thanks.

Check out trac, Charlie has been working on OpenCL detection on the Mac, have a look for mention of a scaling factor for AMD GPUs, where the reported GPU RAM amount is doubled because of a Bug,
this scaling factor keeps getting removed then added again, you can try running the <coproc_debug>1</coproc_debug> logging flag in a cc_config.xml to see if you get any more info,
then report your findings to the Boinc_alpha mailing list, you'll get a quicker response,

Edit: the last change is [trac]changeset:25019[/trac]

Claggy

I asked Rom about this a couple of days ago, here's what he said to me:

The bug was in the ATI drivers for a better part of two years, hence BOINC's weak attempt to fix it by doubling the size reported by OpenCL. When they fixed the bug, they only fixed it for some cards, not all of them... So now we have some funky results.

The bug is deeper down than just the OpenCL stuff. So just checking the OpenCL version isn't enough. I don't yet know what the magic combination of things are to check. So I'm just advocating reporting what the APIs report and let AMD fix the stuff they let slide for so long.

The only affected GPUs so far seem to be the OpenCL 1.0 ones, or the 4xx0 range.

---

Thanks very much to both of you for the responses. It seems that the older 4xx0 series GPUs are going to be problematic for testing new OpenCL apps so I'm going to pick a hopefully more compliant candidate for future experiments. I have access to some 2010 and 2011 iMacs and I'm sure I'll find one of those with a GPU that identifies correctly and has OpenCL 1.1 (or better) rather than 1.0.

Unfortunately, I really don't have time to get involved with testing BOINC. I don't really have much time for testing new Einstein stuff, for that matter. The Albert test project has been around for quite a while and I'm only now able to look at setting up a test machine.

Once again, thanks for the helpful replies.