Sad Milestone: 1200 (now 1800+) days since last prime

[Citrix]

Originally Posted by opyrt

I didn't test that, as we are at n=15M+ now.

I believe it is how much CPU that is used to feed the GPU.

I meant testing 15-16 M or 15M to 15.1 M

[Ungelovende]

Originally Posted by Citrix

I meant testing 15-16 M or 15M to 15.1 M

That was my first test Take a look at the file attatched. 437 candidates from 15M

Speeds:
ppsieve: GTX570 - 225M p/sec with 2 clients running (75% load)
sr2sieve: slow laptop cpu - 26M p/sec 1 client running

[Citrix]

I think this will be useful to avoid some PRP tests.
I think ppsieve can be made faster since the PSP k's are low weight and we can add SPH to it also.
What do you guys think?

[VictordeHolland]

I thought PPsieve was only efficient for big k ranges (like PPS on PG)?

Something to consider though, a GPU uses 100Watt or more while sieving. A SB/IV bridge/ Haswell also uses something like 100Watt. Those CPUs can complete 4 first pass tests <2 days. For the PPsieve to be efficient in terms of eliminated factors/used power the GPU has to eliminate at least 2 candidates <2 days (and thus save 2 FP tests and 2 DC tests).

[Citrix]

Originally Posted by VictordeHolland

I thought PPsieve was only efficient for big k ranges (like PPS on PG)?

Something to consider though, a GPU uses 100Watt or more while sieving. A SB/IV bridge/ Haswell also uses something like 100Watt. Those CPUs can complete 4 first pass tests <2 days. For the PPsieve to be efficient in terms of eliminated factors/used power the GPU has to eliminate at least 2 candidates <2 days (and thus save 2 FP tests and 2 DC tests).

We might be able to achieve this with PPsieve...if we include low weight properties of PSP ks.

[opyrt]

Originally Posted by Citrix

I think this will be useful to avoid some PRP tests.
I think ppsieve can be made faster since the PSP k's are low weight and we can add SPH to it also.
What do you guys think?

To be honest, I have no idea. Based on my own tests, I would say that sieving is not efficirnt at this point. But obviously, all that changes if someone writes a ppsieve that works better in our scenario or if someone successfully ports sr2sieve to GPU (with improved speed).

-Kai

[ltd]

As an info. I had to fight with my migraine this weekend.
Its much better now but I want to take it easy.
So all the things I had promised for this weekend have to be moved a little bit.

Sorry for the delay.

[Citrix]

Originally Posted by opyrt

To be honest, I have no idea. Based on my own tests, I would say that sieving is not efficirnt at this point. But obviously, all that changes if someone writes a ppsieve that works better in our scenario or if someone successfully ports sr2sieve to GPU (with improved speed).

-Kai

I am working on this. I know how to do this mathematically but not sure how to program a GPU (or test it..since I do not have one).

FYI: I did some theoretical testing... if we use the SPH algorithm and only test a range of 500K (n's) (e.g. 15-15.5 M) and only test 33% of all the primes in a range, we can be as fast or may be even faster than sr2sieve (in terms of rate of factors found; when comparing to sieving a 50M dat). This will definitely save some PRP tests.


If some one is good at programming I am willing to help them implement SPH for PSP. It can be done both on GPU or CPU.

[Citrix]

I finally was able to get access to a GPU. Does anyone have a copy of ppsieve CUDA for windows to play with?

[Ungelovende]

Originally Posted by Citrix

I finally was able to get access to a GPU. Does anyone have a copy of ppsieve CUDA for windows to play with?

https://sites.google.com/site/kenscode/prime-programs

[Citrix]

Is the PSP's stats website down? Could some one else confirm. I am having issues getting the webpage to load.