Very interesting, thank you :)

Thank you!!

Thank you for sharing more info Stefan!

Yes, thank you, much appreciated.

Practically we are trying to teach a neural network to calculate molecular energies and forces.

As you well know, the new Nvidia cards are said to be designed for "deep learning". Maybe that will be of some use to you someday.

Yes, these quantum chemistry potential networks are a deep learning application :) So we train them on our local NVIDIA GPUs. But for the moment I don't see a need to distribute the training of the networks to GPUGRID if that's what you meant. It trains fast enough locally.

OK, that gives me a better idea of how we fit into the overall scheme of things. I am glad we can do that work for you.

Thank you for answering my questions Stefan.

I've been running the GPU tasks on a 3570k (along side some other projects) on Ubuntu 14.04, just 4 threads so no chance of the simultaneous start issue. All good so far.

So we train them on our local NVIDIA GPUs. But for the moment I don't see a need to distribute the training of the networks to GPUGRID if that's what you meant. It trains fast enough locally.

I cannot understand.
Are we crunching GpuGrid QM to "prepare" data for your local/internal GPU?

Hm ok, maybe you are not familiar with machine learning. Sorry if I glossed over it. In machine learning and specifically supervised learning as in this project you "teach" a network to replicate some ground-truth calculations (in this case QM energy/force calculations).

This means that we take some molecules, position their atoms in 3D space and you guys calculate with QM the energy and forces of this conformation of this molecule.

Then I locally on my computer show a network only the positions of the atoms in space and ask it to predict the energy and forces that you calculated for us (from the QM). This might sound pointless because why are you predicting stuff you already know? Well the great thing about networks is that they are very good interpolators, so if I now give it a molecule (more or less) similar to the ones it was trained on and ask it what is the energy of this molecule, the network will give me an incredibly good estimate of the energy/forces in a few microseconds while with QM I might need minutes to do the same.

Does this clarify it?

Thank you Stefan, this is a fantastic explanation!

+1

Does this clarify it?

Very, very clear.
Thank you!!

For some reason, that approach appeals to me a lot as an ideal distributed computing project. It allows you to offload the slow calculations, while you retain the flexibility of asking a lot of different questions on the returned data that you can investigate with your high-speed calculations. It does not appear to require excessive bandwidth to shuttle the data back and forth (though I have a lot if needed), nor does it require a lot of memory (though I have plenty of that too if you need it).

Also, it does not appear to be tied to a particular type of molecule or disease, but should lend itself to a wide range of subjects. And lastly, aside from a few startup glitches, it seems to be reliable on home computers, and we need not constantly fight to keep it running in the face of errors or crashes. I hope, and expect, that you will achieve significant success with it.

I am running QC on my box currently and there does not appear any startup glitches. It can run two and it can start multiple at the same time. I see there is some usage of flock in the app, and that seems to solve the startup crashes that everyone is complaining. So, Good job.
Pity that there are no more amd gpu tasks, after I acquired two such gpus.

---

No Chems for me;)
The GPU-WU runs best under my Win7 64 bit-System (Xeon e3-1230 v2 + NVIDIA GTX 1050 +8 Gig RAM).
So, over VMWare 14 Ubuntu 18.04 with newest boinc-client (4 cores, 4 Gig ram, 100 gig HD-Space). No Quantum Chemistry available, he says. Solutions?
Sorry, long night, no patience. Say me i have forget some infos.

So, over VMWare 14 Ubuntu 18.04 with newest boinc-client (4 cores, 4 Gig ram, 100 gig HD-Space). No Quantum Chemistry available, he says. Solutions?

When I started processing QC WUs, I had to configure GPUGrid preferences to use CPU and accept QC tasks.
Then, I had to upgrade several of my Linux sytems from 4 to 8 GB RAM.
4 GB Systems didn't get QC WUs. This seems to be your problem.
And I also had to upgrade two systems with 64 GB SSDs to bigger sizes because a lack of HD space.
You're wellcome. Good Luck!

mmmh, i think i can upgrade ram to 16 GB. CPU, QC allowed. You think, the ram is the problem... not the virtualmachine self?

For sure, Quantum Chemistry WUs won't be sent to 4 GB RAM systems.
Also, BOINC Client must "think" VirtualBox is installed.
BOINC Manager log should show similar to last line in this example, coming from one of my systems running QC WUs successfully...

Sun 03 Feb 2019 19:37:31 WET | | Starting BOINC client version 7.6.31 for x86_64-pc-linux-gnu
Sun 03 Feb 2019 19:37:31 WET | | log flags: file_xfer, sched_ops, task
Sun 03 Feb 2019 19:37:31 WET | | Libraries: libcurl/7.47.0 OpenSSL/1.0.2g zlib/1.2.8 libidn/1.32 librtmp/2.3
Sun 03 Feb 2019 19:37:31 WET | | Data directory: /var/lib/boinc-client
Sun 03 Feb 2019 19:37:31 WET | | CUDA: NVIDIA GPU 0: GeForce GTX 1050 Ti (driver version 415.27, CUDA version 10.0, compute capability 6.1, 4032MB, 3613MB available, 2138 GFLOPS peak)
Sun 03 Feb 2019 19:37:31 WET | | OpenCL: NVIDIA GPU 0: GeForce GTX 1050 Ti (driver version 415.27, device version OpenCL 1.2 CUDA, 4032MB, 3613MB available, 2138 GFLOPS peak)
Sun 03 Feb 2019 19:37:32 WET | | Host name: ServicEnginIC
Sun 03 Feb 2019 19:37:32 WET | | Processor: 4 GenuineIntel Intel(R) Core(TM)2 Quad CPU Q9550 @ 2.83GHz [Family 6 Model 23 Stepping 10]
Sun 03 Feb 2019 19:37:32 WET | | Processor features: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts acpi mmx fxsr sse sse2 ss ht tm pbe syscall nx lm constant_tsc arch_perfmon pebs bts rep_good nopl cpuid aperfmperf pni dtes64 monitor ds_cpl vmx smx est tm2 ssse3 cx16 xtpr pdcm sse4_1 xsave lahf_lm pti tpr_shadow vnmi flexpriority dtherm
Sun 03 Feb 2019 19:37:32 WET | | OS: Linux: 4.15.0-43-generic
Sun 03 Feb 2019 19:37:32 WET | | Memory: 7.79 GB physical, 16.00 GB virtual
Sun 03 Feb 2019 19:37:32 WET | | Disk: 424.24 GB total, 256.27 GB free
Sun 03 Feb 2019 19:37:32 WET | | Local time is UTC +0 hours
Sun 03 Feb 2019 19:37:32 WET | | VirtualBox version: 5.1.38_Ubuntur122592
.
.