Message boards :
Number crunching :
Intel GPU Validation Inconclusive
Message board moderation
Author | Message |
---|---|
Saicere Send message Joined: 9 Jul 99 Posts: 7 Credit: 6,717,907 RAC: 0 |
I recently added two new Skylake 6700K-based computers with the Intel GPU enabled and available for Seti@Home to use, but I'm seeing a lot of "Validation Inconclusive" results on both for work units using the Intel OpenCL application (specifically, SETI@home v8 v8.00 (opencl_intel_gpu_sah) windows_intelx86). https://setiathome.berkeley.edu/results.php?hostid=8004096 https://setiathome.berkeley.edu/results.php?hostid=8004128 CPU-generated units seem fine, and most of the GPU results get updated to "Valid", there's just one "Invalid" at this point in time. Any idea what might be causing the discrepancy and large number of inconclusive results? |
Saicere Send message Joined: 9 Jul 99 Posts: 7 Credit: 6,717,907 RAC: 0 |
Was looking a bit more at this, specifically the one work unit that was marked as invalid: https://setiathome.berkeley.edu/workunit.php?wuid=2160985056 The results look similar, while the AMD GPU and CPU results found 9 pulses and a triplet, the Intel GPU result found 10 pulses. Comparing the actual results from the AMD and Intel GPU, sorted by time with the differing entry at the bottom: AMD: Pulse: peak=0.1869025, time=45.82, period=0.06417, d_freq=1763327896.77, score=1.011, chirp=-20.16, fft_len=128 Pulse: peak=0.5658, time=45.82, period=0.4011, d_freq=1763332430.22, score=1.053, chirp=48.532, fft_len=256 Pulse: peak=6.95606, time=45.84, period=15.58, d_freq=1763333460.1, score=1.052, chirp=-97.251, fft_len=512 Pulse: peak=7.482055, time=45.84, period=15.58, d_freq=1763333451.55, score=1.131, chirp=-97.437, fft_len=512 Pulse: peak=7.584933, time=45.99, period=22.91, d_freq=1763331222.83, score=1.02, chirp=51.262, fft_len=4k Pulse: peak=4.384547, time=45.99, period=10.5, d_freq=1763326967.09, score=1.035, chirp=55.625, fft_len=4k Pulse: peak=5.635067, time=45.9, period=14.58, d_freq=1763325916.18, score=1.032, chirp=-68.506, fft_len=2k Pulse: peak=4.664864, time=45.9, period=10.77, d_freq=1763329940.72, score=1.073, chirp=-76.392, fft_len=2k Pulse: peak=7.535296, time=46.17, period=21, d_freq=1763323584.46, score=1.044, chirp=46.654, fft_len=8k Triplet: peak=10.01984, time=48.65, period=28.95, d_freq=1763323381.63, chirp=43.492, fft_len=512 Intel: Pulse: peak=0.1866636, time=45.82, period=0.06417, d_freq=1763327896.77, score=1.01, chirp=-20.16, fft_len=128 Pulse: peak=0.5739036, time=45.82, period=0.4011, d_freq=1763332430.22, score=1.068, chirp=48.532, fft_len=256 Pulse: peak=6.865847, time=45.84, period=15.58, d_freq=1763333460.1, score=1.038, chirp=-97.251, fft_len=512 Pulse: peak=7.336316, time=45.84, period=15.58, d_freq=1763333451.55, score=1.109, chirp=-97.437, fft_len=512 Pulse: peak=7.647188, time=45.99, period=22.91, d_freq=1763331222.83, score=1.028, chirp=51.262, fft_len=4k Pulse: peak=4.426685, time=45.99, period=10.5, d_freq=1763326967.09, score=1.045, chirp=55.625, fft_len=4k Pulse: peak=5.647213, time=45.9, period=14.58, d_freq=1763325916.18, score=1.034, chirp=-68.506, fft_len=2k Pulse: peak=4.581229, time=45.9, period=10.77, d_freq=1763329940.72, score=1.054, chirp=-76.392, fft_len=2k Pulse: peak=7.537786, time=46.17, period=21, d_freq=1763323584.46, score=1.044, chirp=46.654, fft_len=8k Pulse: peak=1.256464, time=45.84, period=1.668, d_freq=1763332922.82, score=1, chirp=54.879, fft_len=512 Outside of the last pulse/triplet and a minor discrepancy in the peaks, the results are matching, with identical time/period/d_freq/chirp/fft_len. |
BetelgeuseFive Send message Joined: 6 Jul 99 Posts: 158 Credit: 17,117,787 RAC: 19 |
I recently added two new Skylake 6700K-based computers with the Intel GPU enabled and available for Seti@Home to use, but I'm seeing a lot of "Validation Inconclusive" results on both for work units using the Intel OpenCL application (specifically, SETI@home v8 v8.00 (opencl_intel_gpu_sah) windows_intelx86). Same problem for me (i5-6500). There seem to be problems with the OpenCL support in the Intel HD 530 driver. This has also been reported on other projects (e.g. Einstein@home). Only project that I found so far that works without problems is Collatz Conjecture, but this is probably because they are not using floating point. Tom |
BetelgeuseFive Send message Joined: 6 Jul 99 Posts: 158 Credit: 17,117,787 RAC: 19 |
Ah, nearly forgot. You should also check this thread: http://setiathome.berkeley.edu/forum_thread.php?id=78613 Tom |
Saicere Send message Joined: 9 Jul 99 Posts: 7 Credit: 6,717,907 RAC: 0 |
Ah, nearly forgot. You should also check this thread: Thanks for the tip. I looked at the thread in question, but I'm not seeing any kind of runtime error or anything else that indicates that the work unit failed in any way, it just seems to be generating results that are slightly different from some other computers. Without knowing the logic behind the result validator/comparator, it seems sufficiently different to get a "Validation Inconclusive" state but still get validated after a third result. For the GPU results for these two computers, they are now sitting at 20 valid, 1 invalid and 25 inconclusive. Out of the valid ones, quite a few were inconclusive, but then got validated for all three tasks, which seems weird. If the first two task results had discrepancies, shouldn't at least one of them be invalid? I guess it's possible that Intel pulled another FDIV bug with the Skylake HD 530 GPU, but there could be something weird going on with the Seti@Home result validator as well. Also, driver versions: OpenCL: Intel GPU 0: Intel(R) HD Graphics 530 (driver version 20.19.15.4377, device version OpenCL 2.0, 13041MB, 13041MB available, 221 GFLOPS peak) OpenCL CPU: Intel(R) Core(TM) i7-6700K CPU @ 4.00GHz (OpenCL driver vendor: Intel(R) Corporation, driver version 5.2.0.10094, device version OpenCL 2.0 (Build 10094)) |
Raistmer Send message Joined: 16 Jun 01 Posts: 6325 Credit: 106,370,077 RAC: 121 |
there could be something weird going on with the Seti@Home result validator as well. Consider simple example. Let say validator consider results as different if difference more than 2. Let result A be 1, result B be 3.1 Apparently they are different for validator. Then result C comes and it's 1.5 What validator should discard then? |
Saicere Send message Joined: 9 Jul 99 Posts: 7 Credit: 6,717,907 RAC: 0 |
there could be something weird going on with the Seti@Home result validator as well. I'm not saying that the validator doesn't have to make choices like this, but when WU results involving the Intel HD 530 produce "Validation Inconclusive" results 50% of the time, there might be something off with the thresholds for flagging it as such. The other possible conclusion is that the HD 530 GPU and/or OpenCL drivers are just bad at math. |
Raistmer Send message Joined: 16 Jun 01 Posts: 6325 Credit: 106,370,077 RAC: 121 |
The other possible conclusion is that the HD 530 GPU and/or OpenCL drivers are just bad at math. Most probably. We had issues with Intel's OpenCL drivers before too. |
Earthcore Send message Joined: 9 Feb 13 Posts: 19 Credit: 2,333,725 RAC: 0 |
I have the same problem with i5-6200U Skylake and Intel HD Graphics 520 OpenCL 2.0 Intel may have a bug in their GPU in terms of floating point processing? |
Earthcore Send message Joined: 9 Feb 13 Posts: 19 Credit: 2,333,725 RAC: 0 |
After upgraded Intel HD Graphics 520 driver today, to 2016-07-25 version 21.20.16.4494, all work-units ends up with Computation Error (i.e. Beräkningsfel). SETI@home 8.12 SETI@home v8 (opencl_intel_gpu_sah) 06mr09af.1650.18882.5.32.83_2 00:00:01 (00:00:01) 84,94 100,000 - 2016-10-04 11:14:45 0,00581C + 1 Intel GPU Beräkningsfel (5,) What's wrong? |
betreger Send message Joined: 29 Jun 99 Posts: 11361 Credit: 29,581,041 RAC: 66 |
What's wrong? IMHO Intel GPU is what's wrong. Many including myself have found them to not be worth the effort. |
Earthcore Send message Joined: 9 Feb 13 Posts: 19 Credit: 2,333,725 RAC: 0 |
OK, but maybe someone in the project should talk with Intel to fix this problem. |
rob smith Send message Joined: 7 Mar 03 Posts: 22219 Credit: 416,307,556 RAC: 380 |
The problem is that the Intel iGPU family, while very common, are very difficult to get working properly. The developers are all volunteers, and have skills that have been developed on the two "mainstream" GPU families (AMD & nVidia), and not the iGPU. Now, if the project could find an iGPU developer interested in providing their services for free I am sure they would be welcomed with open arms. Bob Smith Member of Seti PIPPS (Pluto is a Planet Protest Society) Somewhere in the (un)known Universe? |
petri33 Send message Joined: 6 Jun 02 Posts: 1668 Credit: 623,086,772 RAC: 156 |
The problem is that the Intel iGPU family, while very common, are very difficult to get working properly. The developers are all volunteers, and have skills that have been developed on the two "mainstream" GPU families (AMD & nVidia), and not the iGPU. I'd hug them open arms. To overcome Heisenbergs: "You can't always get what you want / but if you try sometimes you just might find / you get what you need." -- Rolling Stones |
Michael H.W. Weber Send message Joined: 13 Aug 05 Posts: 3 Credit: 3,250,984 RAC: 0 |
Well, it seems, my Broadwell Intel IGP (HD 5500) has the same issues. Question 1: Wouldn't it be an idea to validate ATI vs. ATI, Intel vs. Intel and NVIDIA vs. NVIDIA GPU tasks? For many projects differences in rounding when using different CPUs result in slightly different results such that a bit-wise result comparison runs into issues when cross-system validation is applied. Question 2: So far, I havn't taken closer look into the details of my tasks, but is it correct that from what was reported above one could come to the conclusion that, in a second round of validation, formerly non-validated (inconclusive) tasks were validated DESPITE THE FACT THAT THE RESULT CONTENT IS NOT IDENTICAL when inspected with bare eye? If that is true, then I would suspect serious problems with alls of these IGP-generated results. Please clarify (I do not know anything about what the basis of the vaildation process is). Michael. |
Richard Haselgrove Send message Joined: 4 Jul 99 Posts: 14653 Credit: 200,643,578 RAC: 874 |
1) SETI doesn't use a bitwise validator: there is a (tight) tolerance allowed on the numeric values. 2) There are two levels of validation: 'stronly similar' and 'weakly similar'. For full validation to take place, at least two of the task results must be 'strongly similar', and one of those will be chosen as the canonical scientific result. If, after validation has taken place, there are other 'weakly similar' results, they are awarded credit as a reward for a 'good try', but the weakly similar results are never used as the scientific outcome. This schema is actually more robust than always comparing 'like with like'. There is still a small possibility that two of the 'weaker' hosts will validate against each other, and their slightly inaccurate data will get adopted as science, but the risk is seen as being low enough to be acceptable within the aims of this project. |
rob smith Send message Joined: 7 Mar 03 Posts: 22219 Credit: 416,307,556 RAC: 380 |
Also we can normally only see the stderr report, which is a summary of the results file (the file that is sent soon after a task has completed, this has a lot more information about what has been found than the stderr file has. Bob Smith Member of Seti PIPPS (Pluto is a Planet Protest Society) Somewhere in the (un)known Universe? |
Michael H.W. Weber Send message Joined: 13 Aug 05 Posts: 3 Credit: 3,250,984 RAC: 0 |
OK. What is your conclusion: Should one make use of the IGPs in Intel CPUs in this project or rather not? Michael. |
TimeLord04 Send message Joined: 9 Mar 06 Posts: 21140 Credit: 33,933,039 RAC: 23 |
OK. The HD 530 on my parents' HP Pavilion, Skylake Desktop system SUCKS!!!!! I will say that it performs SLIGHTLY better on v8 MB Tasks than it did on v7 work. It FAILED MISERABLY on AP Units. The card has HIGH Inconclusive rates; and even WITH Lunatics 0.44 installed, it is SLOWER at processing work than MANY STOCK CPU crunchers. If I had to pay for an Intel GPU; I WOULDN'T DO IT!!! It's unfortunately what came with my parents' system. If you want a REAL GPU, get EVGA NVIDIA or AMD/ATI GPUs. My two main crunchers are using EVGA. In Andromeda, (MAC OS X El Capitan), I'm using TWO GTX-750TI SC cards, and in Exeter, (Win XP Pro x64), I'm using a GTX-760. These two machines run FAR, FAR better than the Skylake system for GPU crunching. TL TimeLord04 Have TARDIS, will travel... Come along K-9! Join Calm Chaos |
rob smith Send message Joined: 7 Mar 03 Posts: 22219 Credit: 416,307,556 RAC: 380 |
Like TimeLord my experience with iGPU is "less than favourable". I know some have made them work fairly well, but I failed. Tasks taking at least as long as on the CPU (an i7) and causing no end of lags and crashes. So I gave up with it. Bob Smith Member of Seti PIPPS (Pluto is a Planet Protest Society) Somewhere in the (un)known Universe? |
©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.