Linux app is live now.

What driver version is needed to use the CUDA8.5 (Linux) app?

8.5 hasn't been released yet so it must be a typo, otherwise the app version would be different too.

nanoprobe reported using the 367.35 driver with his GTX 1060 on Ubuntu 16.04 LTS.
I've just installed 370.28 in the hope it will work with a 1060-3GB tomorrow...
OT I'm seeing more GPU utilization variation with the 370 driver than with my previous 361.42 driver on my 970 (85 to 92% GPU utilization).

NB: 375.10 Beta (Linux) brings support for the GeForce GTX 1050 & GeForce GTX 1050 Ti. On Windows 7/8.1/10 (x64) you will need the 375.63 or newer driver if you have a GTX1050 or 1050Ti (according to NV; though I wouldn't be surprised if slightly older drivers worked fine for the GPUGrid apps).

---

FAQ's

HOW TO:
- Opt out of Beta Tests
- Ask for Help

8.5 was a typo. For Linux, cc>=6.0 and driver 367+ will get the 8.0 app

Matt, What about mixed generation setups? Say a GTX1060 and a GTX980 in the same rig for example? How are apps/tasks allocated in such a setup?

---

FAQ's

HOW TO:
- Opt out of Beta Tests
- Ask for Help

Matt, What about mixed generation setups? Say a GTX1060 and a GTX980 in the same rig for example? How are apps/tasks allocated in such a setup?

With extreme difficulty. If you look at the coproc data transmitted to the server in the sched_request_www.gpugrid.net.xml file, you won't find any mention of the GTX980 card. And if the server doesn't know it exists, then it can't make any tailored allocations to suit the lower spec card.

I was present in Budapest in September 2014 to hear David Anderson's keynote talk A Brief History of BOINC. As slide 52 of workshop_14.pdf indicates, the 'Coprocessor model' is one of the "things we need to change" - this is what they mean. But they haven't done it yet.

The GPUGrid apps read the GPU's directly, so in theory they can check if they should be running on said GPU (Pascal or earlier) or if a different app should be used (cuda 6.5). Just asking if that (or something similar) was done for the cuda 8.0app.

---

FAQ's

HOW TO:
- Opt out of Beta Tests
- Ask for Help

The GPUGrid apps read the GPU's directly, so in theory they can check if they should be running on said GPU (Pascal or earlier) or if a different app should be used (cuda 6.5). Just asking if that (or something similar) was done for the cuda 8.0app.

Unless the app can be supplied as some sort of 'fat binary' combining Cuda 8.0 and Cuda 6.5 paths, or as a separate 'launcher plus two alternative child apps', that doesn't get us any further forward.

Even given that the 8.0 app knows it is commencing to run on a card which "should" have been given a 6.5 app instead (which in effect it is doing now, by crashing), the BOINC framework provides no mechanism for transferring the task, either to run on a different card with the same app, or on the same card with a different app.

As the app's 'Pascal only'; Matt, I'll asume mixed NV generation GPU setups will Not work!
An app_config based GPU/'queue' exclusion won't work because the cuda 8.0 app's populated all queues & app names are <name>Long runs (8-12 hours on fastest card)</name> and so on, they don't specify the CUDA version.

---

FAQ's

HOW TO:
- Opt out of Beta Tests
- Ask for Help

As the app's 'Pascal only'; Matt, I'll asume mixed NV generation GPU setups will Not work!
An app_config based GPU/'queue' exclusion won't work because the cuda 8.0 app's populated all queues & app names are <name>Long runs (8-12 hours on fastest card)</name> and so on, they don't specify the CUDA version.

I don't have any mixed Pascal/Maxwell hosts, but I do have three GTX 970/750Ti combos. The 750Ti cards are a bit weak for the normal long queue cards, so I exclude GPUGrid from them completely, and let another project make use of them.

I'm having a big problem on Pascal work units, every few hours I come back to find my 1080 idling. Long and short work units, no particular doctorate's work either. I have to suspend then resume for it to start again

I'm having a big problem on Pascal work units, every few hours I come back to find my 1080 idling. Long and short work units, no particular doctorate's work either. I have to suspend then resume for it to start again

So I am not the only one with this problem. Good to know.

Haven't seen that on Linux and I don't know what else you're doing on the systems or what your Boinc & System settings are so I can't add much, other than point out that you're both using 375.70.

---

FAQ's

HOW TO:
- Opt out of Beta Tests
- Ask for Help

Haven't seen that on Linux and I don't know what else you're doing on the systems or what your Boinc & System settings are so I can't add much, other than point out that you're both using 375.70.

I updated my drivers to 375.7 because of this problem. There is no difference.

Linux app is live now.

Please add the SWAN_SYNC=1 option to the new Linux app, as there is a significant performance loss without it on the Pascal cards. The proof of it that there is no GTX 1080 has beaten my GTX 980Ti's yet on the performance chart, not even my own GTX 1080 (which runs under Linux now).

Haven't seen that on Linux and I don't know what else you're doing on the systems or what your Boinc & System settings are so I can't add much, other than point out that you're both using 375.70.

I updated my drivers to 375.7 because of this problem. There is no difference.

This could be an app issue but I cant help you with that. I'm suggesting things that you can check/try. In the past there have been power issues (downclocking) that were corrected with new drivers and sometimes worked around using tweaks.

Suggest you check your systems Power settings (sleep/suspend, hybrid sleep, suspend pcie, screen saver). Also, what are your resources settings (Device Profile settings) at WCG? I know you can set them to the following:
Power Saving: Set my profile so that my computer's power settings can take effect
Minimum Impact: Set my profile so that it will have a minimal amount of impact on my computer.
If you change profile settings there, they apply here too (assuming your device is on the same profile), but if you have any local Boinc settings these will override your project profile settings.

---

FAQ's

HOW TO:
- Opt out of Beta Tests
- Ask for Help

I have changed GPU to 560 Ti. Same driver, OS, BM and everything. The problem is gone. So I think that there is a problem with pascal app...

Hi,

My conf:
Ubuntu 16.04
kernel: Linux 4.4.0-47-generic
CPU: Intel(R) Core(TM) i3-4170 CPU @ 3.70GHz [Family 6 Model 60 Stepping 3]
GPU: NVIDIA GeForce GTX 1060 3GB (3012MB)
NVIDIA Driver Version: 370.28
Version de BOINC 7.6.31
Mémoire 15996.81 MB

Well, It's working well, lot GPU, few CPU. Working maybe to well
GPU (nvidia-smi outputs):
# gpu pid type sm mem enc dec command
# Idx # C/G % % % % name
0 1028 C 93 31 0 0 acemd.914-80.bi

# gpu pwr temp sm mem enc dec mclk pclk
# Idx W C % % % % MHz MHz
0 94 60 93 31 0 0 3802 1949

CPU (4 cores):
%Cpu(s): 1,0 ut, 0,5 sy, 5,1 ni, 92,7 id, 0,6 wa, 0,0 hi, 0,1 si, 0,0 st

PID UTIL. PR NI VIRT RES SHR S %CPU %MEM TEMPS+ COM.
2730 boinc 30 10 12,873g 340636 149216 S 24,9 2,1 1:34.18 acemd.914-80.bin


It is possible to limit a little the GPU usage ? (sorry, newbie on boinc)
Because my Firefox who use GPU acceleration have some difficulties...

But nice job !

Linux app is live now.

Please add the SWAN_SYNC=1 option to the new Linux app, as there is a significant performance loss without it on the Pascal cards. The proof of it that there is no GTX 1080 has beaten my GTX 980Ti's yet on the performance chart, not even my own GTX 1080 (which runs under Linux now).

After a couple of workunits processed, I've observed that my GTX 1080 (without SWAN_SYNC) could be up to 7% faster than my GTX 980Ti depending on the batch, but it's usually slower.
Still I think there's a significant performance loss without SWAN_SYNC under Linux, and I can demonstrate this on a PABLO_SH2TRIPEP_L_TRI workunit:
Task 15543899 12.016 sec GTX 980Ti @1367MHz _ _ _ _ _ WinXP, SWAN_SYNC on
Task 15569769 10.528 sec TITAN X (Pascal) @1531MHz (?), Linux, SWAN_SYNC off (87.6% of the GTX 980Ti, i.e. 14.1% faster)
However, it should be much faster:
Let's compare the theoretical computing indices of the two GPUs:

GTX 980Ti:        2816 CUDA cores * 1367MHz = 3.849.472
TITAN X (Pascal): 3584 CUDA cores * 1531MHz = 5.487.104
                      5.487.104 / 3.849.472 = 1.425

-> 42.5% faster i.e. the runtime should be only 70.16% of the GTX980Ti's (8.430 sec)
If I assume that the Pascal is boosting to its max (~2000MHz) than it's even worse:

GTX 980Ti:        2816 CUDA cores * 1367MHz = 3.849.472
TITAN X (Pascal): 3584 CUDA cores * 2000MHz = 7.168.000
                      7.168.000 / 3.849.472 = 1.862

-> 86.2% faster i.e. the runtime should be only 53.7% of the GTX 980Ti's (6.453 sec)

_cufft64_80.dll is downloading at 50% after 5h00 !
What is happening ?
Can somwone post me a link to the file ?

_cufft64_80.dll is downloading at 50% after 5h00 !
What is happening ?
Can somwone post me a link to the file ?

Still downlaoding after an extra night. Basically I cannot crunch GPUGRID tasks at the moment.

_cufft64_80.dll is downloading at 50% after 5h00 !
What is happening ?
Can somwone post me a link to the file ?

http://www.gpugrid.net/download/cufft64_80.dll

---

1 Corinthians 9:16 "For though I preach the gospel, I have nothing to glory of: for necessity is laid upon me; yea, woe is unto me, if I preach not the gospel!"
Ephesians 6:18-20, please ;-)
http://tbc-pa.org