_mat_

Sorry for the delay, just came back from my vacation. Errors like these indicate a malformed file or an errornous upload. The submission files are very sensitive because of their encryption. As the browser switch solved your problem I would guess a browser extension got in the way.

Awesome score! Congrats, Roman.

Not true. GPUPI is just not memory bandwidth limited especially on GPUs where memory is already very fast. It's like SuperPi 1M if you would like to compare them. But you shouldn't because these two benches are very different in its core. GPUPI has its own tweaks in hardware and software but like for every benchmark out there it depends on the bottlenecks to get the most efficieny. But I get what you mean. GPUPI relies on another API+driver layer, like 3D benchmarks with DX and GPU driver. That adds another level of tweaking as well that should be used of course.

Parallel computations like in GPUPI can not be validated during a run. The result is an accumulation of billions of smaller partial results, that have been crunched by each core.

havli, as far as I can see it threads and blocks are not the problem here. In the legacy version for CUDA 6.0 my own code decides the occupancy of the GPU, because there is no kernel analyzer function in this version of the CUDA toolkit. Starting with 6.5 such a function was introuduced in CUDA and therefor it is used in GPUPI. The problem here is more likely some bug in the CUDA implementation. Like some really small issue with a double precision function. ls, as said in my pn this is related to stability of your system and not a bug.

Thanks for your report! Try 1M and let me know if it's working. The result is invalid because the kernels can not be executed anymore after a certain point. I might have to look at this problem closer. Btw, GPUPI 1.4 uses CUDA 6.5, GPUPI 2.1 uses 6.0. I thought I could improve backwards compatibiltiy with 6.0 by supporting Compute Capabilty 1.2. But these old versions won't allow double precison anyway. Well, I might switch back to CUDA 6.5 for the legacy version, which should fix this problem too.

Hello fellow overclockers! I took us some time but last Monday we finally delivered the donations to St. Anna Kinderspital! Turrican's family was with us to hand over no less than 1200 Euro! I wanted to let you know, that they are very grateful for all you guys showing your love and respect for him. Thanks for each and every donation, it wouldn't have been possible without you! You can see Turrican's family on the right side, his parents and his two lovely sisters. Btw, his father looks almost exactly like Karl. We published a little news (in German), you can find it here: https://www.overclockers.at/news/spendenuebergabe-im-andenken-an-unseren-turrican

I've just updated GPUPI to version 2.1.2. It's a minor release that fixes the detection of AMDs Furys and R9 300 graphics cards. If you have any problems with the release, just let me know. Btw, this is not a mandatory release, all versions >= 2.1 will work for submission. Downloads: GPUPI 2.1.2 (1.06 MB, CRC-32: CB9069F2) | GPUPI 2.1.2 - Legacy Version (Windows XP, GeForce 200 series, 590 KB, CRC-32: 1497D087)

So close to beating the 3 minutes with 6 cores! Keep pushing!

Nice efficiency! Great work!

Awesome frequencies! Congrats!

sofos is back! Does the high uncore help the efficiency?

6G! Amazing work!

Thanks for your feedback. Seems like the NVIDIA drivers make no difference betweeen GTX 560-448 and a normal one. But the number of Compute Units are different, so I can fix this on my side. It will be included in my next release. HD 6870 lacks double precision support indeed. No chance to run these cards with any version of the benchmark, sry.

Have a look at the best 100M scores, most of them go for the Intel drivers, like here.You have to know that 100M uses the kernel with only 64 bit arithmetic, while 1B uses 64 bit for partial results up to and including 500M, but a special 128 bit arithmetic kernel for above. That's where the AMD drivers are strong. I would not go for the APP SDK because the drivers are too old. With the 3.0 beta you will only get the Omega 14.2 Catalyst driver. Newer versions of the OpenCL driver are currently only available with the Catalyst for graphics cards.

For 100M on Intel CPUs the Intel OpenCL driver seems to be the fastest choice in my tests. For 1B I would plug in an AMD card and install the newest Catalyst driver and use their OpenCL platform.

Yes, it would be possible to implement a version running on any platform that supports CUDA and/or OpenCL, but there are no plans yet. I am currently focusing my time on GPUPI 2.2 which will bring some interesting improvements. Can't promise anything right now.

I was waiting for something like this. Awesome!

New day, new world record in the 8 core category! Great work!

Perfect example for good tweaking! Your good memory reduction time got you the world record. Congrats!

This is not a 24/7 hotline, it's a support thread. I check it about 5 to 6 times a day, but sometimes I am sleeping. About your problem: After a valid run goto to "Submit to HWBOT" in the menu. A new dialog will open and allow you to "Save to file".

Yes, that's perfectly normal behaviour. It's only possible with 100M and a batch size of 100M. The reason for this is that only 4 batches are sent to the CPU to calculate the result, so the benchmark can't split it up into the usual 20 steps. Thanks Massman! I will introduce a progress bar for submission in the next version, that will make slower uploads more visual. We can also add the mainboard vendor, mainboard name, amount of memory and so. There is still some potential to improve the integration.

I've finished and uploaded GPUPI 2.1.1 which introduces serveral smaller bugfixes: Detection of operating system is no longer including language dependend suffixes. (thx to Taloken & COMIAS) Submission to HWBOT can now be started with an opened result popup "Pi calculation done" (thx to der8auer) Fixed hardware detection for Core i7 980X/990X/995X and QX9750/QX9770/QX9775. Please test it and let me know if that fixes all current issues. Download: GPUPI 2.1.1

rtsurfer, newlife: It's impossible that the other cards are helping in any way if they are not part of the calculation. There is a slim chance for a bug (there always is), but I doubt it because it's a very basic functionality. It's much more likely to be something else like a tweak or a very good combination of parameters and driver, I honestly don't know. The moderation should ask for more proof like a video or an explanation. Roman, thanks for your feedback. Very much appreciated! It's fine to close the popup before submission, I actually implemented it this way. But you are right, people would expect it to work the same way as in SuperPi where the popup had to be part of the screenshot. I will add the possibility to submit with an opened popup in the next version. COMIAS, your exotic Windows versions are the problem indeed. "Microsoft Windows*7 Édition Intégrale" ... This will be fixed with GPUPI 2.1.1, a minor bugfix release coming in the next few days (as soon as possible).

What makes you think that they are helping? If you selected one card out of multiple only that one will calculate the score. The others won't do a thing.