Forum > GPU crunching
Latest drivers (NVidia and ATI)
Jason G:
--- Quote from: Ghost on 05 Sep 2011, 04:44:42 pm ---Is that like ATI saying they will resolve the high CPU usage in Cat 11.8 & SDK2.5 ::) :D
--- End quote ---
Pretty much. That statement would imply they will rewind 2 years of OpenCL (& GPU compute in general) development. I'd say more likely that's some default template response to complaints. GPU supercompute & the technologies they rely on are changing too quickly to maintain outdated methods.
The way to 'fix' the new drivers, bearing in mind they aren't really 'broken' as such (at least not in that respect), would be to provide better / clearer documentation & api for synchronisation, which of course requires application modifications. Maybe they'll cock-up some legacy mode or some such, but the old methods of blocking synchronisation are simply outdated, and fundamentally incompatible with future tech... [Too many interrupts in too short of a time frame, and getting worse with every newer faster model, amplified by packing more cards into the same machine. Desktop multithreaded OSes are designed to work with time slices around 1ms... not fast enough for 1 card using blocking syncs & small kernels even! ]
This Includes the blocking syncronisation used in current Cuda applications, and the modifications needed moving forward are the subject of heavy research by many, e.g:
http://www.gputechconf.com/content/GTC/posters/2010/P01-GPU-to-CPU-Callbacks.pdf
OpenCL 1.1 has some of that stuff already.... I could use it on Cuda now, if it existed... I'm hoping it will appear in Cuda 4.1 (or else I'll have to roll my own using OS primitives)
Jason
Richard Haselgrove:
--- Quote from: Ghost on 05 Sep 2011, 04:44:42 pm ---I've just tested this out of idle curiosity, and with Boinc installed in protected execution mode, and Cat 11.8 SDK2.5 installed, Boinc mngr can't connect to the client let alone see the GPU
I'd love to be able to run Boinc as a service, but think there's a way to go before it gets there
--- End quote ---
If you read Jorden's latest addition to the BOINC thread, it seems that the Manager can't connect to the Client because the client isn't running. And the client doesn't run because of some problem between boincsvcctrl.exe and the new ATI software (I won't categorise it as a 'driver', because that would have no purpose living in the same space as a service control manager). In any event, that's well outside our sphere of interest - Jord has forwarded it to the BOINC developers.
benool:
After some time running Catalyst 11.8 and now going back to 11.6 I can confirm that the 11.8 driver are facing performance issues with OpenCL (at least with my 6670).
With 11.8 I would get constant and continuous 100% CPU usage (of one core) as soon as a GPU task was running.
With 11.6 I get the described 80% of one core usage with times where the CPU usage is much lower.
arkayn:
285.27 has been released in beta
http://www.nvidia.com/object/win7-winvista-64bit-285.27-beta-driver.html
Frizz:
--- Quote from: benool on 07 Sep 2011, 12:42:35 pm ---With 11.8 I would get constant and continuous 100% CPU usage (of one core) as soon as a GPU task was running.
--- End quote ---
This seems to be fixed with 11.9RC. But not for multi-GPU setups. ... %$!?@# AMD/ATI ::)
Navigation
[0] Message Index
[#] Next page
[*] Previous page
Go to full version