Message boards :
Number crunching :
How many tasks can a GPU do at once?
Message board moderation
Author | Message |
---|---|
Super Nova Nerd Send message Joined: 12 Feb 16 Posts: 73 Credit: 2,915,787 RAC: 0 |
I have an MSI R9 270, overclocked from 955 to 1100 on the GPU and the memory OC's from 1400 to 1500. I let it run one task at a time to test and it was stable for 2 full days with no errors. I then upped it to 2 tasks and it has been like that for about 5 days with zero problems. I was wanting to see if I could try to up that to 3. I know there is a point of diminishing returns and this is not exactly a cutting-edge GPU, but I want to see how far I can push it. |
spitfire_mk_2 Send message Joined: 14 Apr 00 Posts: 563 Credit: 27,306,885 RAC: 0 |
Use GPU-z to monitor how busy your card is. Look at GPU Load graph. Most people try to keep it at 90-95%. |
rob smith Send message Joined: 7 Mar 03 Posts: 22222 Credit: 416,307,556 RAC: 380 |
There are two things to consider. First is GPU loading, this will increase to just below 100%, and then sit there as the number of tasks increases. Second is task throughput (the number of tasks completed in a given time period) this will improve as the number of concurrent tasks increases to a point and will then drop rapidly. For most GPUs in most systems the sweet-spot for throughput is either two or three. However the number at which GPU usage maximizes is probably one or two higher than that. Needless to say the "type" of task being done will have an influence, so let the system stabilise for several days (a week at least) before making each step, this will allow a "representative" selection of tasks and not just hit a pile of one type which may skew the data. As far as I'm aware the reason being on most GPUs the bottleneck isn't the processing but that data transfer within the GPU and between the GPU and the host. It will be interesting to see how newer APIs like "Vulkan" and compute technologies like "SoG" have on these seet-spots. Bob Smith Member of Seti PIPPS (Pluto is a Planet Protest Society) Somewhere in the (un)known Universe? |
Mike Send message Joined: 17 Feb 01 Posts: 34258 Credit: 79,922,639 RAC: 80 |
3 instances on a 270 is too much. On my 380 its only a small benefit. So i stick with 2 instances for now. With each crime and every kindness we birth our future. |
iwajabitw Send message Joined: 18 Mar 00 Posts: 19 Credit: 9,285,966 RAC: 0 |
Running 3 on a GTX 980. |
Cruncher-American Send message Joined: 25 Mar 02 Posts: 1513 Credit: 370,893,186 RAC: 340 |
Running 3 on a GTX 980. On my 2 systems, both of which have dual GTX 980s. I find that running 3/GPU gives slightly more RAC than 2. It's hard to tell how much because of the varying character of the workloads (I do v8 and AP on mine), but barely 10% better, if that. My CPUs are i7-4790K and i7-4820K, so not underpowered there, either. |
Mike Send message Joined: 17 Feb 01 Posts: 34258 Credit: 79,922,639 RAC: 80 |
Running 3 on a GTX 980. 980 is high end GPU. 270 is mid range at best. You can`t compare those. With each crime and every kindness we birth our future. |
Joe Januzzi Send message Joined: 13 Apr 03 Posts: 54 Credit: 307,134,110 RAC: 492 |
Both my systems run v8 and AP. Doing Cuda or SoG v8 can have a impact of Wu's per card, depending on CPU. First System i7-3770K with a mix of 4 GPUs (not using IGPU). CPU only feeds GPUs. Note: I run 3 @ a time, either running Cuda or SoG. On this system, SoG wins easily:-) GTX 980 3 @ a time. GTX 980 3 @ a time. GTX 780 3 @ a time. GTX 960 3 @ a time. Total of 12 Wu's crunching. Running SoG at times, adds high CPU usage. At brief times my CPU is at 100% (starving GPUs?). It's mostly between 40-90% depending on AR. On this system it was my GTX 960 keeping me from trying more Wu's @ a time, doing Cuda. Now doing SoG I add the lack of CPU cores to the mix. Once my RAC stops climbing, I might pull the 960 out and try more Wu's at a time on the other GPUs. Time will tell. If we could control Wu's per card, the 960 would be doing 2. Second System Intel Q9550 with 1 GPU. GTX 560 Ti 2 @ a time. (Cuda) CPU 1 @ a time. Total of 3 wu's crunching. Once RAC stops climbing, I might try 2 CPU wu's. Then try SoG. I also run AP's when we get them. Usually when I run them, it's only 1 Wu per system, keeping the crunching total the same. YMMV Joe Real Join Date: Joe Januzzi (ID 253343) 29 Sep 1999, 22:30:36 UTC Try to learn something new everyday. |
Super Nova Nerd Send message Joined: 12 Feb 16 Posts: 73 Credit: 2,915,787 RAC: 0 |
Thank You. I think I will stick with 2. This card is no 980 for sure. With the OC I have it is more in the 760 ti range. I may just get another of the same cards and run 2 on that system. Then 4 would not be a problem. My system runs 2 GPU and 8 CPU tasks at a time. The CPU sits at 100% and the GPU at about 95%. |
Darth Beaver Send message Joined: 20 Aug 99 Posts: 6728 Credit: 21,443,075 RAC: 3 |
what CPU do you have ? Remember to leave 2 cores free if doing 2 AP's on the GPU and at least 1 core free for MB's |
Super Nova Nerd Send message Joined: 12 Feb 16 Posts: 73 Credit: 2,915,787 RAC: 0 |
I have an FX 8350 running at 4.7ghz. I wasn't worried about AP units because I never seem to get them on this one. Very few anyway. Maybe I have it set up wrong. I'm very new to this. I have the GPU set to .5 and .1 CPU for each task in the app config file and I have the max seti can use set to 100% CPU on the computing preferences. Here is my app_config file- <app_config> <app> <name>setiathome_v8</name> <gpu_versions> <gpu_usage>0.50</gpu_usage> <cpu_usage>0.1</cpu_usage> </gpu_versions> </app> <app> <name>astropulse_v7</name> <gpu_versions> <gpu_usage>0.50</gpu_usage> <cpu_usage>0.1</cpu_usage> </gpu_versions> </app> </app_config> |
Cruncher-American Send message Joined: 25 Mar 02 Posts: 1513 Credit: 370,893,186 RAC: 340 |
IIRC the FX-8350, while having 8 cores, has only 4 Floating Point units, so you might want to not be running so many threads (WUs) simultaneously. Does anyone out there have a feel - or some data - that would either make me wrong, or quantify the use of cores vs. FPUs on this CPU? |
Super Nova Nerd Send message Joined: 12 Feb 16 Posts: 73 Credit: 2,915,787 RAC: 0 |
The 8350 has four modules with two distinct processors in each module that can work on two threads simultaneously. Yes there is only one Floating point scheduler, but it has two 128 bit FPU processing units so it can execute two Floating point operations at the same time. |
HAL9000 Send message Joined: 11 Sep 99 Posts: 6534 Credit: 196,805,888 RAC: 57 |
I have an FX 8350 running at 4.7ghz. If you are only running tasks on the GPU the value for CPU_usage doesn't really matter. However if you are also running tasks on the CPU it's generally a good idea to allocate more CPU for the GPU app(s). Otherwise the GPU app(s) may run slower due to lack of sufficient CPU resources. Ideally you would want to use: <gpu_usage>0.5</gpu_usage> <cpu_usage>1.0</cpu_usage> However you may be able to get away with: <gpu_usage>0.5</gpu_usage> <cpu_usage>0.5</cpu_usage> SETI@home classic workunits: 93,865 CPU time: 863,447 hours Join the [url=http://tinyurl.com/8y46zvu]BP6/VP6 User Group[ |
Super Nova Nerd Send message Joined: 12 Feb 16 Posts: 73 Credit: 2,915,787 RAC: 0 |
Thanks Hal. I was kind of thinking the same thing with the .5. I don't think I need to go to 1 with this GPU anyway. I just noticed I have 2 AP units running. Neither on GPU. I have 2 GPU tasks and the 2 AP units running on CPU only. I have never noticed that before. When they started it was showing 3 days for time remaining, but looks like they will be take about 8 hours. One is almost done. |
TBar Send message Joined: 22 May 99 Posts: 5204 Credit: 840,779,836 RAC: 2,768 |
If you check your results you should see the AMD Multibeams require More CPU than the AMD APs. Assuming you only need to free CPU cores for APs is Not correct when using ATI/AMD cards. You need to have free cores for All GPU tasks with AMD cards, more so with the MBs than APs. Back with APv6 the type of blanking used would require almost a complete CPU core at times. That changed with APv7 which happened quite a while ago. Ever since APv7 the AMD MBs require more CPU on AMD cards. I would suggest freeing at least one CPU core per GPU task, especially on an AMD CPU. |
Super Nova Nerd Send message Joined: 12 Feb 16 Posts: 73 Credit: 2,915,787 RAC: 0 |
I am very new to this and know very little about it. What is multibeam? How do I know if a task was multibeam? AP is easy. It says it is ;) All I see is this- SETI@home v8 Anonymous platform (ATI GPU) This- SETI@home v8 Anonymous platform (CPU) or this- AstroPulse v7 Anonymous platform (ATI GPU) |
Zalster Send message Joined: 27 May 99 Posts: 5517 Credit: 528,817,460 RAC: 242 |
There are 2 work units at Seti. AP which you know because it says so. MB are all the others, seti@home v 7 and v 8 |
Brent Norman Send message Joined: 1 Dec 99 Posts: 2786 Credit: 685,657,289 RAC: 835 |
Anything that says seti is a multibeam task. |
TBar Send message Joined: 22 May 99 Posts: 5204 Credit: 840,779,836 RAC: 2,768 |
Look here, http://setiathome.berkeley.edu/sah_status.html Splitter status: Progress Multibeam Astropulse There are only Two types of tasks, therefore, if it's not an AP it has to be a MB. SETI@home is just another name for MultiBeam. |
©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.