Proth primes with k = 1281979 - Page 3

bur · 2021-12-02, 17:33

Thanks, I had thought about using the GPU since at Primegrid sieving Proth numbers with GPU is much faster than any CPU. The reason I didn't was that I only have a GTX 760 and a GTX 1660. The 760 is quite slow and not useful. The 1660 might be ok, but I prefer to use it for Wieferich/Wall-Sun-Sun search currently and as Happy said, it'd need to be really fast to compete against the 12 cores of the Ryzen 9 3900X.

Btw, hijack all you want, I'm glad to hear about such things.

dannyridel · 2021-12-04, 06:45

Quote:

Originally Posted by bur

Thanks, I had thought about using the GPU since at Primegrid sieving Proth numbers with GPU is much faster than any CPU. The reason I didn't was that I only have a GTX 760 and a GTX 1660. The 760 is quite slow and not useful. The 1660 might be ok, but I prefer to use it for Wieferich/Wall-Sun-Sun search currently and as Happy said, it'd need to be really fast to compete against the 12 cores of the Ryzen 9 3900X.

Btw, hijack all you want, I'm glad to hear about such things.

I agree, just using 6 cores on my 3800xt provides around 4-5 mp /s but using -G 2 - g40 (which seems to be the limit of improving efficiency for me) on a gtx 1650 with ddr6 only gives me around 1 mp/s.

bur · 2021-12-06, 09:53

While we're on the topic of sieving on GPU, did anyone try colab sessions for it? I don't have any experience with it, just began copy&pasting GPU72 code which seems to run fine. Is there a similar "fire&forget" available for srsieve2cl?

bur · 2022-01-19, 10:36

No new primes, just another status update:

No sieving was done since the last update.

All n < 5,600,000 have now been checked. No prime since more than 5M candidates, low weight indeed. :)

Since the FFT size grew to 640K with n > 5.6M, the 64 MB L3 cache of the Ryzen 9 3900x ran out when testing 12 numbers simultaneously. Initially I ran six 2-threaded LLR instances, but noticed that two of them were about 30% slower than the other four. The reason being the special layout of the processor. There are four so-called CCDs with 16MB L3 cache each. And since each CCD houses three cores, that means that two of the LLR instances ran on two separate CCDs.

So I switched to four 3-threaded LLR instances occupying a single CCD each. Maybe special constructs like 4 2-threaded and 4 single-threaded LLRs would lead to a higher throughput, I didn't run any tests.

Smallest LLR-test currently running:
n = 5.62M
FFT = 640K
duration = 4060 s / test
digits = 1.69M
Caldwell entry rank: 241

Largest LLR-test currently running:
n = 5.65M
FFT = 640k
duration = 4090 s / test
digits = 1.70M
Caldwell entry rank: 238

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Proth primes	ET_	Proth Prime Search	14	2021-12-16 02:08
Proth 2.0	pepi37	Software	10	2020-09-06 17:26
Proth and Riesel Primes	lukerichards	Number Theory Discussion Group	7	2018-01-20 16:47
(NEW) Proth Primes Section	kar_bon	Riesel Prime Data Collecting (k*2^n-1)	6	2010-11-25 13:39
some primes I found with Proth.exe last year	ixfd64	Lounge	1	2005-09-07 23:42

2021-12-02, 17:33	#23
bur Aug 2020 796581e-4;32539e-3 1032₈ Posts	Thanks, I had thought about using the GPU since at Primegrid sieving Proth numbers with GPU is much faster than any CPU. The reason I didn't was that I only have a GTX 760 and a GTX 1660. The 760 is quite slow and not useful. The 1660 might be ok, but I prefer to use it for Wieferich/Wall-Sun-Sun search currently and as Happy said, it'd need to be really fast to compete against the 12 cores of the Ryzen 9 3900X. Btw, hijack all you want, I'm glad to hear about such things.

2021-12-06, 09:53	#25
bur Aug 2020 796581e-4;32539e-3 538₁₀ Posts	While we're on the topic of sieving on GPU, did anyone try colab sessions for it? I don't have any experience with it, just began copy&pasting GPU72 code which seems to run fine. Is there a similar "fire&forget" available for srsieve2cl?

2022-01-19, 10:36	#26
bur Aug 2020 796581e-4;32539e-3 2×269 Posts	No new primes, just another status update: No sieving was done since the last update. All n < 5,600,000 have now been checked. No prime since more than 5M candidates, low weight indeed. :) Since the FFT size grew to 640K with n > 5.6M, the 64 MB L3 cache of the Ryzen 9 3900x ran out when testing 12 numbers simultaneously. Initially I ran six 2-threaded LLR instances, but noticed that two of them were about 30% slower than the other four. The reason being the special layout of the processor. There are four so-called CCDs with 16MB L3 cache each. And since each CCD houses three cores, that means that two of the LLR instances ran on two separate CCDs. So I switched to four 3-threaded LLR instances occupying a single CCD each. Maybe special constructs like 4 2-threaded and 4 single-threaded LLRs would lead to a higher throughput, I didn't run any tests. Smallest LLR-test currently running: n = 5.62M FFT = 640K duration = 4060 s / test digits = 1.69M Caldwell entry rank: 241 Largest LLR-test currently running: n = 5.65M FFT = 640k duration = 4090 s / test digits = 1.70M Caldwell entry rank: 238