what? you don't have to PAY for CUDA. it's a platform, and the cuda toolkit is free from nvidia. they just need a developer who knows it or can learn it. anyone can code with CUDA. when SETI was still live, we had a volunteer who wrote his own application basically from scratch, in CUDA, and provided it absolutely free. I even helped compile it at times and I have absolutely no affiliation with nvidia or SETI other than being a user. that app was 4-5x faster than the SETI provided OpenCL app, partly due to CUDA optimization, and partly due to the great search algorithms and coding techniques by the author.

From what I understand, nvidia provided some resources to SETI to help develop one of their earlier CUDA apps, but I'm not aware that SETI actually paid for that. I got the impression it was more of a PR/marketing campaign from nvidia to promote GPU computing.

CUDA is just faster. on similarly spec'd hardware CUDA on NVIDIA will trump OpenCL on AMD.

While, I don't have any experience myself with CUDA programming, to me it seems only fair to point out the obvious speedup of the GPU app over the CPU app, but on the same time the obvious inefficiencies as well that still exist with this app. I noted this same observations at the beginning of March numerous times along with others during the beta tests, when the first proposal to run several WUs concurrently was initially made to increase util. However, I still see 2 problems with this approach. First, I agree with Ian&Steve on their stance that CPU resources are essentially wasted. If the GPU were more demanding/would require more computational power, less WUs would have to be run concurrently to bring up GPU util and thus free CPU threads. Second, with the current WU supply, the app_configs specifically finetuned to allow running of multiple WUs concurrently on a card, is at least for now never completely satisfied and thus restricts performance of the few tasks that actually get downloaded and run on the card. In my case allowing each card to run 3 WUs but only getting two, means that 1/3rd of the card's computational power cannot be used.

Is this app open-source for users to review? Or only the research results?

Hope that some smart people can contribute beyond their hardware to enhance efficiency of the app even further.

Hi there,

i got a bunch of GPU-WUs, unfortunately all failed as in BETA.

Here are some:
https://www.worldcommunitygrid.org/ms/device/...s.do?workunitId=609761634
https://www.worldcommunitygrid.org/ms/device/...s.do?workunitId=609761638
https://www.worldcommunitygrid.org/ms/device/...s.do?workunitId=609761636
https://www.worldcommunitygrid.org/ms/device/...s.do?workunitId=609761586
...more...

Basically they all failed the same way.
System: Windows 10 Pro, Q9550S, 8GB RAM, NVidia GT640 with OpenCL 1.2 driver installed.
[edit] GPUz says it is GK107 revision A2 chip (Kepler) and is capable for OpenCL 1.2 [/edit]

Other projects, f.e. milkyway run without problems.

For now i have disabled GPU for WCG until further notice. :(

[edit2] A native CUDA application would be an absolute killer! ;) [/edit2]

I still don't know what you mean by paying for "cuda support". in no way will the researchers have to pay money directly to nvidia for this. they simply need a developer with the skills or ability to learn it. having the ability to write a cuda application is not something that nvidia holds a ransom for.

if you mean by "support" as paying for the developer's time, then it's no different than paying for openCL support. you're just paying the person to do the work.

having openCL is nice for compatibility (nvidia and intel and AMD GPUs with basically one source, just compiled differently) and it's a little easier for the development team, but CUDA would speed up processing on nvidia cards. I'm fairly sure the researchers couldn't care less about what apps or devices things are run on, they just want the results as fast as possible. I'm not proposing to abandon OpenCL completely, keep it for the Intel and AMD GPUs that need it (they don't have their own special platform), but it would be nice to have a CUDA app for nvidia since it's faster. though not many projects supply both a CUDA and openCL app probably for dev/time reasons. they usually pick one or the other.

even better!