New CUDA application for Windows x64 Release
Message boards :
News :
New CUDA application for Windows x64 Release
Message board moderation
Author | Message |
---|---|
Send message Joined: 9 Jun 12 Posts: 584 Credit: 52,667,664 RAC: 0 |
Last modified: 23 Mar 2020, 9:23:07 UTC We are proud to present you the new CUDA application that was released by our team yesterday. What's improved? - We've Added support for latest Compute Capabilities (CC). Application was built with latest CUDA SDK v10.2 adding support for latest NVIDIA devices with Compute Capabilities (CC). This build supports CC 3.0, 3.5, 3.7, 5.0, 5.2, 6.0, 6.1, 7.0, & 7.5 from Kepler, Maxwell, Pascal, Volta & Turing architectures and their GeForce, Quadro, NVS and Tesla series, but does NOT supports Tegra and Jetson SoC Series devices (CC 3.2, 5.3, 6.2 & 7.2) which are not subject of interest for this project. - Fixed issue with special Light curves which has more than 1000 points. There is no x86 (32bit) version. CUDA 10.2 supports x64 (64bits) operating systems only. Side effects that could be observed because of the update. If your client shows unusual "Remaining time" for the tasks you can follow these steps: 1) Set "No new tasks" from "Project" tab for "Asteroids@Home" project. 2) Then you need to finish your recently downloaded tasks. 3) When all are done and reported to the server please reset the project with 'Reset project' button at Project tab. We also suggest to use the latest stable version 7.14.0 of BOINC client. And last but not least we strongly suggest users, who has CUDA enabled devices to upgrade their drivers to the latest one. For today it is 442.74 - WHQL. You will find it here: https://www.nvidia.com/Download/driverResults.aspx/158759/en-us For legacy drivers you will find drivers here: https://www.geforce.com/drivers/beta-legacy Despite new CUDA application is release we will continue to support older devices with older drivers by the previous CUDA55 application. I would also like to thank Georgi Vidinski who made this release possible. Radim Vančo (FoxKyong) |
Send message Joined: 30 Oct 19 Posts: 2 Credit: 726,372 RAC: 287 |
|
Send message Joined: 22 Nov 17 Posts: 159 Credit: 13,180,518 RAC: 0 |
Hi mg13 [HWU], Thank you for participating in Asteroids@home. Creating OpenCL application to support AMD GPUs was always at our RoadMap for years. Even now there is ongoing development on that. No mater there were some turbulence during the years, postponing and getting back on, we are doing our best to make it happen. Once we have solid PoC application we will be glad to provide it to our contributors. Regards, Georgi. |
Send message Joined: 1 Jan 14 Posts: 302 Credit: 32,671,868 RAC: 2 |
Hi mg13 [HWU], Perhaps since Seti is closing you can get a copy of their version to cut the process in half. |
Send message Joined: 14 Dec 14 Posts: 1 Credit: 3,708,480 RAC: 0 |
|
Send message Joined: 8 Dec 12 Posts: 8 Credit: 5,631,804 RAC: 5 |
Last modified: 24 Mar 2020, 5:34:38 UTC I noticed something else that looks strange. I am running the new application on a computer with 2 graphics cards (GTX 1070 Ti and GTX 2070). In BOINC manager it shows 2 GPU tasks running (one on device0 and one on device1 as expected), but when I look at the GPU load using GPU-Z, I see that the GPU Load is 98% for the GTX 2070, but always 0% for the GTX 1070 Ti. The workunits are completing successfully, but it doesn't look like the GTX 1070 Ti is actually being used. I am using Windows 10 Professional with latest updates, BOINC Manager 7.14.2 and NVidia driver v445.75 I have tried resetting the project, but this didn't help. Edit: I just tried it on another computer with 2 GPUs (two identical GTX 760 cards) and the same thing is happening. Even though BOINC manager shows 2 GPU tasks running, I only see load on 1 GPU. |
Send message Joined: 22 Nov 17 Posts: 159 Credit: 13,180,518 RAC: 0 |
Is it just me or is v102 SLOWER than v55 ... if it is slower ... what is the scientific advantage to this version? Hi mjfrybar@yahoo.com, Thank you for participating in Asteroids@home! There could be several reasons for the application being a bit slower than previous version. External factors like newly added Work Units from our new sources has different impact on overall calculation process despite their size. Could be anything from number of Light curves, number of their points, how data is structured. These are all factors on which we don't have control as data is not always, let's say "linear" structured, as scientists may wish. It's mater of series of observation specific events at the source of the data. There have been a few changes inside the code too. There was an issue which caused computation errors on random tasks because of a misaligned synchronisation of code. It was observed on very rare occasions and still we had to fix that. After all speaking of and doing science is not matter of speed. It is matter of precision. Thanks. Georgi |
Send message Joined: 15 Mar 20 Posts: 4 Credit: 361,920 RAC: 0 |
|
Send message Joined: 22 Nov 17 Posts: 159 Credit: 13,180,518 RAC: 0 |
I noticed something else that looks strange. Hi Daniel, Thank you for participating in Asteroids@home and for your feedback. May I ask you for two more things to check about the issue you are pointing? Can you please take one note on the execution time per task when both cards are running tasks and another one when just single card is running. Because of the nature of the application there should be significant difference in execution times if by some reason it has happened two task being running on a single GPU card. Can you also check what Sensor screen of HWInfo shows in both cases as well? The way BOINC client addresses GPU cards is very straight forward using "deviceId"-s which should guarantee that different GPU cards are addressed at the same time. Thanks. Georgi |
Send message Joined: 22 Nov 17 Posts: 159 Credit: 13,180,518 RAC: 0 |
The new CUDA102WIN hangs after few fractions of an percent,... Hi HausGeist, Thank you for participating in Asteroids@home. Actually application did not hangs at all. The gaps between reported fractions done is because during that time the whole computation is done on the Device (GPU card) and the Host (Computer) does not knows how far it went. So don't worry, everything is fine and works as expected. You can check that by the status of finished tasks of yours. We are aware of that side effect and will try to improve the reporting routine in feature release. Thanks, Georgi |
Send message Joined: 19 Nov 14 Posts: 2 Credit: 100,565,760 RAC: 0 |
|
Send message Joined: 8 Dec 12 Posts: 8 Credit: 5,631,804 RAC: 5 |
Last modified: 24 Mar 2020, 10:26:37 UTC Hi Georgi, Thanks for your reply. I've done some more quick tests and found the following ... With 2 tasks running HWiNFO64 says ... GTX 1070 Ti "GPU Core Load" = 0% GTX 2070 "GPU Core Load" = 98% Runtime is around 28 minutes. With 1 task running HWiNFO64 says ... GTX 1070 Ti "GPU Core Load" = 0% GTX 2070 "GPU Core Load" = 100% Runtime is around 13 minutes. |
Send message Joined: 22 Nov 17 Posts: 159 Credit: 13,180,518 RAC: 0 |
Hi Daniel, Thank you for the information. That helps a lot. That means those two task are working on same GPU card definitely. Which is very interesting. We'll need to investigate that issue deeper as it turns out BOINC client does not addresses both GPU cards correctly but reports that tasks are running on separate deviceIds in the same time. Can you check one more thing please. Which card what deviceId has assigned to it, like: GTX 1070 Ti | "GPU Core Load" = 0% | DeviceId = x GTX 2070 | "GPU Core Load" = 100% | DeviceId = y It is interesting to see what's going on. Meanwhile you can limit the number of simultaneous running tasks on GPU to one for the Asteroid@home project. Sorry for the inconvenience. Thanks. Georgi |
Send message Joined: 30 Oct 19 Posts: 2 Credit: 726,372 RAC: 287 |
Hi mg13 [HWU], Thank you so much for the answer Georgi. Do you have an estimate of when this application will be available? I also wanted to ask you, why don't you see my avatar well? |
Send message Joined: 9 May 13 Posts: 6 Credit: 8,918,481 RAC: 2,875 |
I noticed something else that looks strange. Same problem here with GTX 970 and 1660 Super. I had the same issue with 2 Radeon at Moo! Wrapper some month ago and never found a solution. |
Send message Joined: 7 Mar 14 Posts: 78 Credit: 6,720,028 RAC: 1,831 |
Last modified: 25 Mar 2020, 6:42:03 UTC Forgive my ignorance, but I have to ask. Is the new CUDA application suitable for normal 64-bit Windows 7 computers or only for certain brands/models? My current Windows 7 computer has an AMD A-6-6400K processor with Radeon HD graphics, 8 GB of RAM and a 1TB hard drive7. It crunches Asteroids, along with SETI@Home and Rosetta@Home, 24/7/365. So is this computer compatible with the new CUDA application? In the instructions for installing this new application it is specified to complete all running tasks, then select "no new tasks" and reset the project. My account still shows four task with validation pending. If my computer does indeed support this new application, should I wait for these tasks to be validated before resetting the project? Thanks. Steven G Oldsmar, Florida |
Send message Joined: 28 Dec 18 Posts: 7 Credit: 963,309 RAC: 0 |
Last modified: 25 Mar 2020, 9:36:02 UTC Forgive my ignorance, but I have to ask. Is the new CUDA application suitable for normal 64-bit Windows 7 computers or only for certain brands/models? CUDA = Nvidia OpenCL = Radeon |
Send message Joined: 22 Nov 17 Posts: 159 Credit: 13,180,518 RAC: 0 |
I'll try to shed some more light regarding new CUDA application and the hardware that it addresses. What is CUDA? CUDA® is a parallel computing platform and programming model developed by NVIDIA for general computing on graphical processing units (GPUs). More info HERE and HERE. That means CUDA applications can only works on NVIDIA based video cards (GPUs) thus will be sent only to BOINC clients with NVIDIA GPUs enabled settings. In our case there are two application versions already. First we had cuda55 which can be run on older models with Compute Capabilities (CC) from 2.0 till 6.5, and the new one, cuda102, which was released on Monday 23rd of March, and which supports as well as older devices, starting with CC 3.0 and the newest too, with CC 7.5. This application does not addresses and can't be run on AMD based Radeon GPUs. Radeon GPUs usually run applications developed on top of the OpenCL libraries which is different from CUDA and thus it is not subject to the announcement. More info HERE and HERE. I hope this was helpful to those, who was confused. Regards! Georgi |
Send message Joined: 11 Aug 13 Posts: 1 Credit: 849,120 RAC: 0 |
|
Send message Joined: 7 Mar 14 Posts: 78 Credit: 6,720,028 RAC: 1,831 |
|
Message boards :
News :
New CUDA application for Windows x64 Release