Message boards :
Number crunching :
Intel 5th gen GPU with SETI v8 issue?
Message board moderation
Author | Message |
---|---|
KillerDemon Send message Joined: 10 Jul 11 Posts: 6 Credit: 282,738 RAC: 0 |
Hi, I am experiencing the following issue: As as soon as I let my 5th gen i5 HD5500 intel GPU run the SETI v8, all my CPU tasks will go very slow and have days remaining instead of only a few hours. My system is Windows 8.1 with the latest Intel Graphics drivers and latest BOINC installed. If I put usable cpu cores at 75% or 50% then the same issue occurs with the remaining tasks when I let the GPU run, very slow CPU tasks. I hope somebody can help because I had to turn off the GPU for now. Ps. on v7 I have not experienced this issue |
Raistmer Send message Joined: 16 Jun 01 Posts: 6325 Credit: 106,370,077 RAC: 121 |
There should be no changes between v7 and v8 in this area. Perhaps app's tuning differs. Try to add -cpu_lock to command line. Before that turn off CPU usage in BOINC completely and look via task manager or ProcessExplorer what CPU usage pattern for GPU app. Does it use different CPUs or only single logical CPU shows some load? And CPU instances slowdown running along with GPU instance on combined device (like AMD's APU or Intel's equivalents) is unavoidable. Both compete for cache and memory bandwidth resources. Look this article as example: http://lunatics.kwsn.info/index.php/topic,1735.0.html |
KillerDemon Send message Joined: 10 Jul 11 Posts: 6 Credit: 282,738 RAC: 0 |
Ok, I have paused all CPU tasks so only the GPU runs. it has avarage 2% usage on SYSTEM, 2% on System Interrupts and 2% on the SETI GPU app. Total 6% usage. With 2x CPU task and 1x GPU task I have 75% CPU usage. However, the CPU tasks now go very slow; 0.001-0.002% per second instead of 0.01-0.015% per second when I suspend the GPU task. I get the feeling that the GPU task has some kind of high priority and gets favored above CPU computation in general. My whole system also feels slow in general when I enable v8 GPU task. :( Edit: I have an i5-5300U with HT enabled (4 logical cores) |
Raistmer Send message Joined: 16 Jun 01 Posts: 6325 Credit: 106,370,077 RAC: 121 |
Ok, I have paused all CPU tasks so only the GPU runs. That's not as interesting as how this % distributed between cores? Perhaps would be easely just to look in GPU app process preperties and see what is affinity for this process? Also, do you see slowdown of all CPU tasks or just single one? |
KillerDemon Send message Joined: 10 Jul 11 Posts: 6 Credit: 282,738 RAC: 0 |
Hi, The affinity is (all) 4 logical core. However changing it to 1 does not affect the problem. All the CPU apps slow down and so does everything else (outside BOINC) that uses CPU. Edit: I have notice the following: With 4x CPU and 1x GPU, the CPU FAN spins at a much slower speed. When I disable GPU, the CPU fan starts spinning up again. CPU Usage remains at 100% in both situations |
Mike Send message Joined: 17 Feb 01 Posts: 34253 Credit: 79,922,639 RAC: 80 |
The only solution to this in my experience is freeing a CPU core and add no_cpu_lock switch. With each crime and every kindness we birth our future. |
Raistmer Send message Joined: 16 Jun 01 Posts: 6325 Credit: 106,370,077 RAC: 121 |
Well, if affinity don't change situation that no need -cpu_lock switch. Seems like this device suffers from hardware design. Maybe because of TDP limitations (that is, some external limitation) maybe just some bus saturation (inner limitations). In APU-like devices CPU and GPU parts are interrelated. So performance of such device breaks superposition law. CPU performance + GPU performance != Device performance. If you have srong will to maximize device performance I highly recommend to conduct same measurements I described in cited article. Then you will find sweet spot for your particular device. For some low-end Intel devices that sweret spot quite possibly will be to disable GPU part crunching completely. BTW, this should be taken into consideration @time of purchase: peak device performance VERY different from peak core performance * number of cores + GPU performance. Even considering only CPU part one can notice highly non-linear performance dependance from number of cores in use. |
Raistmer Send message Joined: 16 Jun 01 Posts: 6325 Credit: 106,370,077 RAC: 121 |
The only solution to this in my experience is freeing a CPU core and add no_cpu_lock switch. Mike, it's iGPU with different defaults. As OP stated affinity non-locked by default there. Freing CPU on such device can degrade overall performance cause on Intel CPU part usually far better than GPU one. |
KillerDemon Send message Joined: 10 Jul 11 Posts: 6 Credit: 282,738 RAC: 0 |
Ok thank you for your reply. The system is my work-owned Latitude E5450 so I cannot make any hardware changes. I will disable the GPU computations for now on this system. Thank you for your time and help! Ps. @ Mike On v7 it indeed ran fine CPU app performance wise when I freed one CPU core. Too bad that doesn't work on v8 for me anymore. |
Raistmer Send message Joined: 16 Jun 01 Posts: 6325 Credit: 106,370,077 RAC: 121 |
That is real puzzle actually cause nothing should be changed in v8 iGPU app that could result in such different behavior... And regarding hardware changes - no, I don't propose any hardware changes. We all play with what we have here :) Just some load-balancing to optimize device throughput. But in first approach just disabling GPU part will be OK perhaps. |
-= Vyper =- Send message Joined: 5 Sep 99 Posts: 1652 Credit: 1,065,191,981 RAC: 2,537 |
Have you tried https://downloadcenter.intel.com/download/24075/Intel-Extreme-Tuning-Utility-Intel-XTU- so it isn't harder on the gpu resulting in TDP max kicking in to slow the cpu/gpu down compared to V7. I have reported earlier issues with my Broadwell and Iris Pro 6200 and it reaches it ceilings directly with AP and sometimes it aborts it due to errors. _________________________________________________________________________ Addicted to SETI crunching! Founder of GPU Users Group |
KillerDemon Send message Joined: 10 Jul 11 Posts: 6 Credit: 282,738 RAC: 0 |
Hi, Thanks, I think it is indeed TDP related. In that tool I see that my CPU core speed goes to ~1.6 GHz when I enable GPU. This happens with 4x, 3x and 2x CPU + GPU. With 1x CPU + GPU the CPU goes to 1.8-2 GHz. With only GPU, core speed goes to normal 2.7 GHz. With 4x CPU, core speed is normal 2.7 GHz as well. Meh. I hope they get AMD APU's in my next machine cycle. |
-= Vyper =- Send message Joined: 5 Sep 99 Posts: 1652 Credit: 1,065,191,981 RAC: 2,537 |
If you want you can adjust so you push the cpu harder. If it's a smalltop, notebook i wouldnt suggest it but if it is a regular computer with big cpu fan etc you can up the limits so you push it harder. Please try to stay below 80 degrees C to be safe. _________________________________________________________________________ Addicted to SETI crunching! Founder of GPU Users Group |
KillerDemon Send message Joined: 10 Jul 11 Posts: 6 Credit: 282,738 RAC: 0 |
If you want you can adjust so you push the cpu harder. If it's a smalltop, notebook i wouldnt suggest it but if it is a regular computer with big cpu fan etc you can up the limits so you push it harder. Normal operation is around 70 degrees when SETI runs, so I think I shouldn't push it higher. :) It is a 14" Latitude E5450 notebook so that would be torture. :) Thanks for the tool anyway, it made clear what the problem is. |
Raistmer Send message Joined: 16 Jun 01 Posts: 6325 Credit: 106,370,077 RAC: 121 |
+1, Thanks, Vyper. Always good to get experimental confirmation to any theories. |
ExtraTerrestrial Apes Send message Joined: 5 Jan 02 Posts: 13 Credit: 1,370,403 RAC: 0 |
Meh. I hope they get AMD APU's in my next machine cycle. THe reason here is neither Intel nor DELL. It's that your CPU has 2 cores with HT, which by itself can saturate the TDP limit. And a GPU which can use most of the TDP for itself. By doing this you get maximum performance if only one of them is working hard. But the price you have to pay is reduced performance under simultaneous load. The only resolutions would be a higher TDP or slower chips (same sustained performance under full load, less performance under partly load). The 1st option exists, but not in Ultrabooks. The 2nd one.. well, you wouldn't gain anything from that. MrS Scanning for our furry friends since Jan 2002 |
Grant (SSSF) Send message Joined: 19 Aug 99 Posts: 13722 Credit: 208,696,464 RAC: 304 |
The only resolutions would be a higher TDP or slower chips 3rd option, discreet graphics. Grant Darwin NT |
Richard Haselgrove Send message Joined: 4 Jul 99 Posts: 14650 Credit: 200,643,578 RAC: 874 |
I've just started pushing the boat out on a new acquisition, to see how it behaves. Gigabyte Z97P-D3 motherboard with Intel i5-4690 CPU @ 3.50GHz. It was assembled by a local trade supplier, in a well-ventilated tower case with a proper CPU cooler (not Intel stock), and sold as a 'Pro gaming' model. GPUs are: NVidia GTX 970 (new supply as part of the package) in x16 slot NVidia GTX 750Ti (scavenged from failed host) in physical x16 slot, wired for x4 Intel HD 4600 I've been running both NVidia cards during the recent testing - usually GPUGrid on the 970, 2xSETI_v8 Cuda50 on the 750Ti. I also run 4xNumberFields@Home tasks - light integer maths, not as optimised as SETI apps - on the four CPU cores. I checked with the Intel Tuning tool before I started, and it was saying 100% CPU utilisation at 3.70 GHz, with a package temperature <=75C and no thermal throttling. The Intel GPU was idling at 600 MHz. Then, I restarted the machine with an HDMI dummy plug in the Intel port. Add to the above one SETI MB iGPU task... Package temperature has risen to <=80C, and Intel GPU speed to 1,200 MHz (varying). Tuning app still reports CPU core speed as 3.70 GHz (which must be turbo mode, since the CPU isn't an overclockable 'K' model). The HD 4600 has reported its first task (4675907728) while I've been typing this - 75 minutes for AR 0.44, including first-time compilations, but no command-line tuning. I think the CPU AVX/64 application would take roughly twice as long for a similar WU. I'm seeing a possible slowdown on the secondary NV GPU (limited memory access, perhaps, at x4), but no other side effects. No screen lag on the HD 4600 (dummy plug - monitor is connected to the GTX 970, and seems to be coping with everything I ask of it). |
-= Vyper =- Send message Joined: 5 Sep 99 Posts: 1652 Credit: 1,065,191,981 RAC: 2,537 |
+1, Thanks, Vyper. No worries! Glad to be to assistance one way or Another. I still got that pesky Astropulse aborting error compared to the V7,V8 apps so i've disabled Astropulse OpenCL for now. _________________________________________________________________________ Addicted to SETI crunching! Founder of GPU Users Group |
©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.