might as well compile it for CUDA 11.8 to bring Ada (40-series) support.
____________

大家好！ 我在中国上海 如何让GPU 工作在百分之一百的状态 我发现在运行时GPU 一直在百分之30左右！[img][/img]

大家好！ 我在中国上海 如何让GPU 工作在百分之一百的状态 我发现在运行时GPU 一直在百分之30左右！[img][/img]

我Nvidia能到80%。我也同时在运行其他的CPU（20%）和Intel GPU（97%）项目。电源调成最佳性能后，CPU到50%。Intel i7 12代。

Looking around I see the present batch of protein ligand sims are crashing... DARNIT!

process exited with code 195 (0xc3, -61)</message>
<stderr_txt>
22:58:08 (3209098): wrapper (7.7.26016): starting
22:58:25 (3209098): wrapper (7.7.26016): starting
22:58:25 (3209098): wrapper: running /bin/bash (run.sh)
/bin/bash: run.sh: No such file or directory
22:58:26 (3209098): /bin/bash exited; CPU time 0.001795
22:58:26 (3209098): app exit status: 0x7f
22:58:26 (3209098): called boinc_finish(195)

Looking around I see the present batch of protein ligand sims are crashing... DARNIT!

process exited with code 195 (0xc3, -61)</message>
<stderr_txt>
22:58:08 (3209098): wrapper (7.7.26016): starting
22:58:25 (3209098): wrapper (7.7.26016): starting
22:58:25 (3209098): wrapper: running /bin/bash (run.sh)
/bin/bash: run.sh: No such file or directory
22:58:26 (3209098): /bin/bash exited; CPU time 0.001795
22:58:26 (3209098): app exit status: 0x7f
22:58:26 (3209098): called boinc_finish(195)

anything else found?

actually I have some on my system so i took a look.

there appear to be many things wrong.

the job.xml file is calling just tar, with no reference to what tar is. this should probably be /bin/tar to use the system tar.

the extracted run.sh script looks woefully lacking in detail. i can see it trying to call python and conda from 'bin/' but that is not included in the input package and will fail. the input tarball only includes some text/config files and not the whole python package.
____________

What app exactly?

Yes, we should have made a beta, but this app is not related to this thread.

interesting to see that Ada "should" run on the Ampere cubins. I know the app has an architecture compatibility check, and it may fail there even if it could otherwise work.

you could also consider compiling your apps with the PTX version for forward compatibility

like this:
-gencode=arch=compute_86,code=sm_86
-gencode=arch=compute_86,code=compute_86

and the user can set the environment variable as needed. or you could set it in the wrapper config file
____________

I am successfully running the current ACEMD_3 tasks on a GTX980ti, on a Quadro P5000, and on two RTX3070.
However, they fail on a GTX1650 after a few seconds:

https://www.gpugrid.net/result.php?resultid=33263379
https://www.gpugrid.net/result.php?resultid=33263343

can anyone tell me what might be the reason?

As a first, you can try resetting GPUGRID project at failing host.
But probably the reason is 4GB RAM being too short for executing these tasks.

...
But probably the reason is 4GB RAM being too short for executing these tasks.

...
But probably the reason is 4GB RAM being too short for executing these tasks.

that's what I am guessing, too.
However, I was closely watching the RAM usage (via MemInfo) when the tasks started: at the moment the task crashed, about 2 GB were still free.
Further, for the tasks running on the other hosts mentioned above, the Windows tasks manager shows a RAM usage between 60MB and 400MB per task.
Maybe the CPU Intel Core2 Duo E7400 @ 2.80GHz is too old for these tasks?
(However, some other GPU projects like Einstein, WCG and Primegrid are running well).

it could very well be that the CPU is too old. it does not support AVX extensions for example, and if the application is built with this requirement then that could be a reason.

Just had one and it failed after 26 seconds on my 4090

Just had one and it failed after 26 seconds on my 4090

Python run fine on my 4090, though they don't do much at all, all the work seems to be on the CPU.

Python run fine on my 4090, though they don't do much at all, all the work seems to be on the CPU.

Done :)

Looks like the application does not understand the 4090 architecture. Needs to be recompiled with the gencodes that Ian pointed out.

ACEMD failed:
Error compiling program: nvrtc: error: invalid value for --gpu-architecture (-arch)

it could very well be that the CPU is too old. it does not support AVX extensions for example, and if the application is built with this requirement then that could be a reason.

perhaps one of the GPUGRID people could tell me if this is the case?

...
But probably the reason is 4GB RAM being too short for executing these tasks.

that's what I am guessing, too.
However, I was closely watching the RAM usage (via MemInfo) when the tasks started: at the moment the task crashed, about 2 GB were still free.
Further, for the tasks running on the other hosts mentioned above, the Windows tasks manager shows a RAM usage between 60MB and 400MB per task.
Maybe the CPU Intel Core2 Duo E7400 @ 2.80GHz is too old for these tasks?
(However, some other GPU projects like Einstein, WCG and Primegrid are running well).

onl(unknown error) - exit code 195 (0xc3)</message>
<stderr_txt>
23:38:59 (9616): wrapper (7.9.26016): starting
23:38:59 (9616): wrapper: running bin/acemd3.exe (--boinc --device 0)
23:39:01 (9616): bin/acemd3.exe exited; CPU time 0.000000
23:39:01 (9616): app exit status: 0xc0000135
23:39:01 (9616): called boinc_finish(195)y this in the siderr

Looks like the application does not understand the 4090 architecture. Needs to be recompiled with the gencodes that Ian pointed out.

ACEMD failed:
Error compiling program: nvrtc: error: invalid value for --gpu-architecture (-arch)

Is ACEMD3 not yet supporting the NV 4k architecture on W10? This is a 4070 Ti with the CUDA 1121 app.

ACEMD failed:
Error compiling program: nvrtc: error: invalid value for --gpu-architecture (-arch)

Is ACEMD3 not yet supporting the NV 4k architecture on W10? This is a 4070 Ti with the CUDA 1121 app.

ACEMD failed:
Error compiling program: nvrtc: error: invalid value for --gpu-architecture (-arch)

Is ACEMD3 not yet supporting the NV 4k architecture on W10? This is a 4070 Ti with the CUDA 1121 app.

ACEMD failed:
Error compiling program: nvrtc: error: invalid value for --gpu-architecture (-arch)

That’s correct. The current CUDA 11.21 app does not support Ada 4000 series.

I got ACEMD 3 task for my gtx 1080ti on Windows (2oiq-ADRIA_KDMD_1k_test_3809-0-1-RND9959).
GPU stays at very low clock speed 750Mhz and VRAM 800Mhz. I expected 2x CPU clock and 7x VRAM clock. Or would it not have any advantage?