A nasty trick

[seba2122]

Hello again!
Here's my idea that came to my head at 3:10 AM.

I have 2,5Ghz processor with 4 cores.
Thus boinc is running 4 tasks. (1 core ~ 1 processor ~ 1 task)

Now here comes my CLEVER idea.
Is there a way... to 'fool'/cheat CPU tasks to run on GPU cores? I know GPU is about "quantity" not "quality", GPU has many cores/processors with slow clock, but because there is A LOT of them, they are fastrer than CPU (for certain tasks).

So let's say I have like 200+ GPU cores.
Because they are mega slow, one task takes 5 hours to complete, yet I can run 200 of them.
my CPU only handles 4 tasks/hour
200tasks/5hours = 40tasks/hour

I realise that some projects are only for CPU and some are only for GPU, but why not try this? any pros/cons?

Waiting for your answers!!
-----------
Edit: I guessed 200 cores. didn't really do a research for other graphic cards but my "GeForce 330M"(which is useless because I can't run GPU projects with it SOMEHOW) has 48 CUDA cores with up to 1,2GHz clock. This is half of my CPU clock but still 48 cores.
If time scales with clock, then 1 task should take 2 hours to complete so with my SICK MATH I should get 24tasks/hour.
still profit but IS IT POSSIBLE??
-----
Edit 2: Ok GeForce GTX TITAN X has 3072 Cuda cores... Well.. Now I'm excited.

[LaurV]

Your trick won't work, because the "things" don't work like that. You can use your GPU cores, but in a different way. Each "core" in our CPU does some funny task called "iterations", and if you have 4 cores, you can do 4 iterations (of 4 different exponents, each exponent one iteration) in the same time. You can also "ask" more cores to work for the "same" iteration, making it faster. The LL test is in such a way that you can not work more iterations of the same exponent in the same time, each iteration depends on the result of the previous, and it can not be started before the previous iteration is finished, but you can work 4 different exponents in your 4 cores, or only 2 of them. i.e. instead of 4 cores doing 4 iterations for 4 different exponents in the same time, you can use the same 4 cores to do 2 iterations in the same time, for 2 exponents, every group of 2 cores working at the same iteration of the same exponent.

You will half the times, like instead of doing 4 expos and finish in 20 days, you will do 2 expos and finish in 10 days. At the end, you do the same work in 20 days, but you can see some results faster if you are curious. You can also work the same exponent with all 4 CPU cores (all cores will work for the same iteration, of the same exponent, quartering the time), and get it done in 5 days. Satisfying your curiosity in this way came with a gain when your computer can not be involved in the project for long times (you just concentrate to finish one exponent faster), or when you have slow memory channels which can not handle the transfers for 4 exponents), but it also came with a penalty, however, because more cores working at the same exponent have to "agree" each other what each one is doing, split the work, collect the results together, wait for each other (nanoscopic times), so at the end, you do a bit more work, you will need 5 days and a little bit, and you will finish 4 exponents, working them one by one, one after the other, in 22 days, instead of 20 (as it would be if you let each core handle its own exponent). Of course, the numbers here are only examples, your mileage will vary according to your hardware.

Coming back to your GPU, because those 2000 cores are "very silly cores", they can not work "by themselves", what you can do with it is to "ask" all cores (well, as many of them as possible) to work for the same iteration, finishing it much faster. About 3-, 6-, 10-times faster than the 4 CPU cores will do (depending on your GPU/CPU combination). You have 500 times more cores in the GPU, but they are 100 times more stupid, and they generally relay on (about the) same memory/channels/architectures/resources (i.e. you have two channels of memory on the CPU, but only 4, 6 or 8 on the GPU, and not 1000), you will not expect to work 500 exponents, or to work one exponent 500 times faster.

But you can use the GPU to do LL test, and have it done very fast, if you have a good GPU card (like for example a Titan). Why do you think we discuss the subject so much here? We not only discuss, but people invested their money and time and knowledge to make software that can do this. If you want to try, for your GPU, look for "cudaLucas". Also, look here, to see your target performance (scroll down, click on your GPU and compare with your CPU).