Nvidia GPU for PRP testing proth numbers?

[Batalov]

Originally Posted by diep

Busy on a riesel sieve in CUDA as we speak :)

status: busy improving its speed.

First incarnations will only sieve a single k. Maybe later on a few more kernels for
sieving multiple k's...

Originally Posted by Angular

Good points although a GPU sieve for specialized prime chains is unlikely as well :(

Writing a chain sieve like NewPGen is what bit(prime)coiners did. That's why there are some truly randomly looking chains of lengths 9..14 in the top
So it is not all that unlikely.

[diep]

Originally Posted by Angular

Nice work and its sounds a bit complex dealing with the NVidia GPU.

Any chance this would work for fixed n, and large k range k * 2^n+-1?

We can have a look to build such kernel.
Yet some time after it works great :)

Comes time comes more options.

> Or a small n range with a large k range of say k ranging 4 billion?

Ah i get impression a number don't understand what i mean with '32 bits'. 32 bits means: "a number between 2 and up to 2^32". Which of course is 4 billion.

That's what is already working. Busy speeding that up. Right now it is speed of 6 newpgen cores with with fixed k and n range of 7 million.

This for a GTX580 that's delivering on paper a whopping amount of instructions a clock.

Hope to have speeded it up a lot by end of day (but don't bet on it).

Extending k to 63 bits number (so up to 2^63) shouldn't be major problem either if that's in high demand. It is some cut and pasting and a few lines extra.

Range of n i use right now to benchmark is n=7 million. That's coincidence though. Could pick tad smaller. Yet a range that much above n=10 million mightl be major pain for the GPU.

>My projects with Cunningham primes are still using Newpgen (485 MB bitmap limit) and I would love to have another option.

(Must admit i don't know, or have forgotten or whatever, what cunningham primes are - yet will figure it out one day (again) i have the feeling)

I compare speed versus newpgen. Yet always the start you'll have to do with newpgen up to above p > 32 bits.

Realize it is just in phase of getting setup this program.

Some things will be fast on GPU, once you want sieve a limited n range and thousands of k's, then realize that basically making a perfect hashtable is a fast way to move forwards. The hashtable size in the L1 datacache (shared memory) of the GPU is very limited.

As soon as you can do things without hashtable, or with a tiny one, and it's about which entity can push through most instructions a clock for an embarrassingly parallel task, then GPU always wins of course.

[henryzz]

Flexibility with things like Cunningham chains, twins(or more generally k-tuples) would always be appreciated where possible.