Re: gpu client
Kevin McCoy <[email protected]>
| Newsgroups | gmane.comp.distributed-net.rc5 |
|---|---|
| Message-ID | <[email protected]> |
I have been overclocking my GeForce 8600 GTS and running the CUDA client for a couple weeks now. The GPU clock is running at 735 MHz. Mem clock is set to 2100 MHz. No smoke or flames. :-) I keep an eye on the GPU temperature and it floats around 57-60C, depending on the ambient temperature. I have a water-cooled Intel Core-2 Quad, overclocked to 2781 MHz in the same case (Gigabyte 3D Mercury) running dnet 24/7/365. CPU temp stays at about 38C with all 4 cores running dnet. The 8600 GTS is just factory stock air-cooled. The only problem with the CUDA client is that it makes the video performance pretty miserable, even with normal applications. I have to pause the client to get any work done. Maybe I need a faster video card and then throttle back the CUDA client :-) SideWinder wrote: On 16.2.2009 16:50, bert wrote: I guess that means I'm going to overclock my card then.....lol Impressive numbers SideWinder! That's probably not advisable, at least unless you have improved cooling. Dnetc will load the GPU much harder in terms of total utilization over time periods measured in several minutes, than any game or benchmark tool for example (excluding stress testers like S&M). So if you OC the card enough to cause any visible artifacts in games, it will likely screw up the RC5 decryption calculation also (at least run it with the "-test" mode command immediately after running normally for a while). Also, expect much reduced electrical life for the card for any serious OC without improved cooling and running this application. Running CUDA is very different than trying to barely get 3DMark to pass with a higher clock and letting it cool between attempts. At least with my current card, the equilibrium temps with Dnetc running are much higher than the average temp during 3DMark Vantage (stock cooling, which is a two slot copper heatpipe HSF). The card is "factory OC'd" by using select silicon, faster spec memory and a custom HSF, it is guaranteed (or more accurately, was guaranteed) to run anything at the factory default clocks, which are higher than nVidia spec for the GPU. I have ruined a few cards with too aggressive overclocking, even without running anything 100% 24/7 over CUDA/Stream. _______________________________________________ rc5 mailing list rc5-Ra3b/QYEcJ3d140v2zMXi0fjHoOT/h/[email protected] http://lists.distributed.net/mailman/listinfo/rc5