Questions and Answers :
GPU applications :
Too much CUDA work!
Message board moderation
Author | Message |
---|---|
Mike Bunce Send message Joined: 30 Jun 99 Posts: 22 Credit: 14,519,146 RAC: 5 |
This post is in two parts: First a question and then my own problem and solution for it. The question: How quickly is an average sized CUDA work unit processed compared with a CPU-processed one? I know that's like asking how long is a piece of string, but I am just looking for ballpark comparisons from people running different GPUs. The reason I ask this is because I have an AMD Phenom 4 core processor that polishes off a work unit per core in about 3-and-a-half hours, but a rather pedestrian GeForce 9400GT that takes 2-and-a-half hours per CUDA unit. From recent posts about bandwidth I get the impression that some GPUs can finish off a CUDA unit in a few minutes! My PC is happily running with all 4 CPU cores crunching SETI 24/7 even when I am computing on it, but I can only use the CUDA when the PC is not in use (ie overnight). My cache of 4 days' CPU work is optimal, but this size of cache results in my getting 50 or 60 CUDA jobs, so I have had to block the download of CUDA to drain the pipeline so to speak. Here is my solution. I stop accepting CUDA in my preferences, then when I need more I set my cache size to 1.5 days and then tick the GPU box. When I get my CUDA quota I untick the GPU box and reset my cache size back to 4 days. This way I can ride out most outages without being in danger of having CUDA units that pass their deadline. I hope this is of use to others with the same problem. Unless you know of a better solution.... |
The Weasel Send message Joined: 6 Jun 99 Posts: 127 Credit: 53,205,208 RAC: 0 |
its all in the cards and your system, a quad core system will put out results but different cards have different results, my 8600 gts card will process a cuda unit in about 30-45min, but my card if overclocked and the cooling on it has been changed. hope that helps |
Dale Mc Send message Joined: 17 May 99 Posts: 2 Credit: 1,191,226 RAC: 0 |
I have a 8600GTS too...how did you change your cooling on the card? I tried Riva, but the fan keeps resetting to a slower speed. How do you stop it, if you use Riva, I get computation errors when it over heats...or is it just Cuba errors ? |
The Weasel Send message Joined: 6 Jun 99 Posts: 127 Credit: 53,205,208 RAC: 0 |
I purchased a Thermatake Sorb cooling unit for my card, and is drawing power from my power supply not from my card. I included the links for the aftermarket cooler: http://www.thermaltake.com/product_info.aspx?PARENT_CID=C_00001166&id=C_00001167&name=SOrb&ovid=n I'm using EVGA Precision v1.7.1 to overclock my card. There is a link somewhere on the seti community pages but I did not write it down. The EVGA download is free, but you do have to create an account (also free) and be logged in under the account you created in order to download the file. Hope this info helps you. |
figment Send message Joined: 1 Aug 03 Posts: 3 Credit: 1,420,981 RAC: 0 |
My rig is a C2D E6700@3.25ghz + MSI GTX 260(factory OC). CPU get about 1.5-2.5 hrs per unit per core. GPU gets some units that process in 3 minutes, and another set that takes about 13 minutes. I'm currently leaving it running 24/7 and suspend it only when playing 3D games.. if I'm just surfing, email, office apps, then I keep it going full throttle. So far, I'm not hitting any deadlines on my LONG list of units. I'm set to accept 2 days of work. ~figment~ |
Mike Bunce Send message Joined: 30 Jun 99 Posts: 22 Credit: 14,519,146 RAC: 5 |
Thanks for that. It's just the sort of comparative I was looking for. Time for me to upgrade my GPU I think! |
figment Send message Joined: 1 Aug 03 Posts: 3 Credit: 1,420,981 RAC: 0 |
Oh, I forgot to mention a little bit of detail.. I'm running lunatics optimized code.. it really makes a big difference in processing time. http://lunatics.kwsn.net/index.php ~figment~ |
©2024 University of California
SETI@home and Astropulse are funded by grants from the National Science Foundation, NASA, and donations from SETI@home volunteers. AstroPulse is funded in part by the NSF through grant AST-0307956.