View Single Post
Old 2021-06-22, 20:32   #4
Mar 2021

59 Posts

Originally Posted by MJansen View Post
Hi Craig,

thanx for the info! I have been reading up on the nvidia guide, but it is a lot to take in and it seems like a tool kit that needs to take into consideration memory management, parallel programming, sending tasks to cores-warp's-SM's, writing your own math functions, retrieving results, preventing dead locks ... pfff no easy task! Cudo's to you for having come this far!

Conceptually I am trying to wrap my head around what can be done in parallel. You wrote that you have each thread (equals 32 cores? So 28 x 128/32 = 112 threads simultaneous? Would be a lot more than the 24 of my AMD 3900 ...) process a single prime (or do you mean prime gap?). Forgive me if I get this wrong, but this confuses me a little.

Regarding a parallel approach, my assumption would be, but I do not know if it is possible using a GPU, that having each of the 112 thread's check a single candidate of a prime gap (with a Fermat test or Miller - Rabin test) and after no PRP's are found, feed the next 112 candidates until a PRP is found, that would be a method. Of course the workload could be smaller, so you can work on more intervals at a time I guess.

Or sieving the interval around the primorial start value, for composites so only the stronger candidates remain, that can be checked using a pseudoprime test on the cpu would be another approach I can think of. But I am totally new to this and nowhere near knowledgeable enough to have a good understanding of the possibilities of using the GPU to its fullest! Again my curiosity speaking: what conceptual approach have you implemented?

Ps no room in my house that needs an extra heater ;-) and 1564 was my mistake, I did scramble the correct core number (3584).

Kind regards
It is actually 3584 cores. I was trying to say that you cannot have 3584 separate codes running at the same time. The cores are broken up into blocks of 32 called warps which all run the same code at the same time. You can have all 3584 cores running a separate Fermat test.
CraigLo is offline   Reply With Quote