Choose your own K and work on finding a top-5000 prime! - Page 33

storm5510 · 2020-06-29, 13:20

Quote:

Originally Posted by pepi37

Proth2.0 run on GPU for base2
https://github.com/galloty/proth20

Quote:

proth20 is a new highly optimised GPU application, created in 2020.

There is a link for an older CPU version, but not the one above.

Viliam Furik · 2020-06-30, 15:42

Hello, I would like to test primes for k=20020913 (my birthday date

). Could somebody please guide me to how far sieve, and what program to use? I have Radeon VII, RTX 2080Ti, and Ryzen 9 3900X available for testing. I would also like to ask if there is any difference between searching in k=20020913 and k=2002913?

In case there isn't a difference, I reserve the k=20020913, and the other one is for the future me.

VBCurtis · 2020-06-30, 16:15

I suggest you make a short project out of n=1 to n=400k to find small primes, and then decide how large an 'n' you want to set up a sieve for.

For the small primes, srsieve or newpgen to create the sieve file, then sr1sieve because it's faster than the other two. One generally sieves until the average time to find a factor is equal to the LLR time for a candidate 70% of the way from n-min to n-max; in this example, that would be at n=280k.
The reason I don't give a sieve depth number is that the proper depth depends quite a bit on how many candidates remain in the sieve, which is a way of saying the weight of the k-value. Different k's have different patterns of small factors that eliminate a bunch of candidates, so some k's need sieve depths an order of magnitude larger than others.

LLR is the primality-testing program; there is a version for GPU, but it's best described as alpha-version quality. Rogue has srsieve2 package of programs, some of which are GPU-enabled; peruse that thread (or wait for someone who knows more than I do here) to see if you can sieve on GPU and test LLR on CPU.

Large sieve ranges are more efficient than small ones; after you run to 300k or 400k for the first sieve and test all those candidates, I'd make a second sieve that covers any range you are likely to ever test- 400k to 4M isn't crazy, but that's core-years of searching; 400k to 2M could be completed in a few months.
The second k will yield different primes, so "saving" that for future-you makes sense.
Happy hunting!

Viliam Furik · 2020-06-30, 18:54

BTW, before any tests, I found 20020913*2^12-1 to be prime. I simply calculated the value and used an online tool (https://primes.utm.edu/curios/includes/primetest.php) to find out whether it is prime.

Happy5214 · 2020-07-01, 08:58

Quote:

Originally Posted by VBCurtis

LLR is the primality-testing program; there is a version for GPU, but it's best described as alpha-version quality. Rogue has srsieve2 package of programs, some of which are GPU-enabled; peruse that thread (or wait for someone who knows more than I do here) to see if you can sieve on GPU and test LLR on CPU.

Mark Rodenkirch (rogue) has stated before that the discrete log algorithm used in srsieve2 (BSGS) doesn't really benefit much from the expanded parallelism provided by AVX or GPUs, so such versions of srsieve2 don't exist.

Viliam Furik · 2020-07-01, 19:06

I meant the GPUs for LLR testing. Is there a program for GPU testing, not sieving? Especially if it had properties as gpuOwl, it would be really fast on Radeon VII.

paulunderwood · 2020-07-01, 19:13

Quote:

Originally Posted by Viliam Furik

I meant the GPUs for LLR testing. Is there a program for GPU testing, not sieving? Especially if it had properties as gpuOwl, it would be really fast on Radeon VII.

I don't think there is any Riesel-GPU application. Why not crunch Mersenne-PRP with gpuOwl on your Radeon VII?

I asked Yves Gallot about Proth on GPUs, downloaded and compiled the source and it was equal to 4 CPU cores, whereas Mersenne is like 40 cores

storm5510 · 2020-07-01, 23:28

Quote:

Originally Posted by paulunderwood

...Why not crunch Mersenne-PRP with gpuOwl on your Radeon VII?...

Unless I am way behind, the only form gpuOwl will accept would be k = 1. There was no way to specify a higher k the last time I looked at it. It would be nice if it would.

Happy5214 · 2020-07-02, 09:18

Quote:

Originally Posted by VBCurtis

LLR is the primality-testing program; there is a version for GPU, but it's best described as alpha-version quality.

Quote:

Originally Posted by paulunderwood

I don't think there is any Riesel-GPU application. Why not crunch Mersenne-PRP with gpuOwl on your Radeon VII?

I asked Yves Gallot about Proth on GPUs, downloaded and compiled the source and it was equal to 4 CPU cores, whereas Mersenne is like 40 cores

https://www.mersenneforum.org/showthread.php?t=14608 describes llrCUDA, which is what I believe VB is referring to. There does not appear to have been any work on it since 2018, and it is definitely alpha-level. As the name suggests, it's CUDA-only.

kuratkull · 2020-07-02, 12:33

llrCUDA works, I have been running one for months. But it's slow, much slower than a CPU in the same class .
I am getting 1.1ms per iteration on n=3.8M on a 1080TI
While the same n is getting 0.271 ms per iteration on 5 threads and 0.7ms on 1 thread on an i7-6700K

Happy5214 · 2020-07-14, 08:05

My work on filling in gaps for (near-)Woodall k's has reached n=300k. Three primes were found:

8508301*2^261263-1
1195203*2^262935-1
8508301*2^296247-1

A prime for k=197673 was found earlier.

The list of k's tested to n=300k is:

197673
251749
582833
586085
665127
938237
1183953
1195203
1268979
1467763
8508301

Updates will be posted to Prime-Wiki shortly.

I'm also reserving k=1993191 (the latest near-Woodall k) for n < 3986382.

2020-06-30, 18:54	#356
Viliam Furik "Viliam Furík" Jul 2018 Martin, Slovakia 1440₈ Posts	First prime, before any tests! BTW, before any tests, I found 20020913*2^12-1 to be prime. I simply calculated the value and used an online tool (https://primes.utm.edu/curios/includes/primetest.php) to find out whether it is prime.

2020-07-02, 12:33	#362
kuratkull Mar 2007 Estonia 225₈ Posts	llrCUDA works, I have been running one for months. But it's slow, much slower than a CPU in the same class . I am getting 1.1ms per iteration on n=3.8M on a 1080TI While the same n is getting 0.271 ms per iteration on 5 threads and 0.7ms on 1 thread on an i7-6700K Last fiddled with by kuratkull on 2020-07-02 at 12:39

2020-07-14, 08:05	#363
Happy5214 "Alexander" Nov 2008 The Alamo City 963₁₀ Posts	My work on filling in gaps for (near-)Woodall k's has reached n=300k. Three primes were found: 85083012^261263-1 11952032^262935-1 85083012^296247-1 A prime for k=197673 was found earlier. The list of k's tested to n=300k is: 197673 251749 582833 586085 665127 938237 1183953 1195203 1268979 1467763 8508301 Updates will be posted to Prime-Wiki shortly. I'm also reserving k=1993191 (the latest near-Woodall k) for n < 3986382. Last fiddled with by Happy5214 on 2020-07-14 at 08:17*

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Best Work for Finding Primes	Unregistered	Information & Answers	9	2012-06-24 13:50
Which settings to choose for best work?	jmb1982	Software	2	2009-04-07 09:33
Help test 210885 - Find a new top 5000 prime!	SlashDude	Riesel Prime Search	121	2008-01-03 08:47
Help test 2995125705 - Find a new top 5000 prime!	SlashDude	Riesel Prime Search	538	2007-05-08 01:42
The fastest way to a top-5000 prime?	lsoule	15k Search	13	2005-09-19 20:24

2020-06-30, 15:42	#354
Viliam Furik "Viliam Furík" Jul 2018 Martin, Slovakia 2⁵×5² Posts	Hello, I would like to test primes for k=20020913 (my birthday date ). Could somebody please guide me to how far sieve, and what program to use? I have Radeon VII, RTX 2080Ti, and Ryzen 9 3900X available for testing. I would also like to ask if there is any difference between searching in k=20020913 and k=2002913? In case there isn't a difference, I reserve the k=20020913, and the other one is for the future me.

2020-06-30, 16:15	#355
VBCurtis "Curtis" Feb 2005 Riverside, CA 13135₈ Posts	I suggest you make a short project out of n=1 to n=400k to find small primes, and then decide how large an 'n' you want to set up a sieve for. For the small primes, srsieve or newpgen to create the sieve file, then sr1sieve because it's faster than the other two. One generally sieves until the average time to find a factor is equal to the LLR time for a candidate 70% of the way from n-min to n-max; in this example, that would be at n=280k. The reason I don't give a sieve depth number is that the proper depth depends quite a bit on how many candidates remain in the sieve, which is a way of saying the weight of the k-value. Different k's have different patterns of small factors that eliminate a bunch of candidates, so some k's need sieve depths an order of magnitude larger than others. LLR is the primality-testing program; there is a version for GPU, but it's best described as alpha-version quality. Rogue has srsieve2 package of programs, some of which are GPU-enabled; peruse that thread (or wait for someone who knows more than I do here) to see if you can sieve on GPU and test LLR on CPU. Large sieve ranges are more efficient than small ones; after you run to 300k or 400k for the first sieve and test all those candidates, I'd make a second sieve that covers any range you are likely to ever test- 400k to 4M isn't crazy, but that's core-years of searching; 400k to 2M could be completed in a few months. The second k will yield different primes, so "saving" that for future-you makes sense. Happy hunting!

2020-07-01, 19:06	#358
Viliam Furik "Viliam Furík" Jul 2018 Martin, Slovakia 1100100000₂ Posts	I meant the GPUs for LLR testing. Is there a program for GPU testing, not sieving? Especially if it had properties as gpuOwl, it would be really fast on Radeon VII.