20200803, 11:27  #562 
"James Heinrich"
May 2004
exNorthern Ontario
3^{2}·337 Posts 
GPUTF threw off the balance of TFPM1. In the old days (preGPU) there was alternation with lowTF, then P1, then highTF, then LL. GPUTF changed the equation by making TF much cheaper. It might be useful to reexamine the methodology with the current set of GPUTF, CPU/GPUPM1, CPU/GPUPRP+Cert software and see if the current order and limits of factoring methods is still optimal.

20200803, 19:02  #563 
"Bill Staffen"
Jan 2013
Pittsburgh, PA, USA
604_{8} Posts 
I'm working from very old and sluggish memories here, but the problem with running p1 on GPUs was having enough memory to set useful bounds, right? That's why we do the gputf; more discrete sizes?

20200803, 23:27  #564  
"TF79LL86GIMPS96gpu17"
Mar 2017
US midwest
2^{2}×1,109 Posts 
Quote:
GPUs go higher on TF because they are much faster at the single precision computations used in TF relative to DPtype computations used in primality testing or P1; ratios ~11 to 40 or so are common. CPUs' speed ratios are much smaller, typically 0.7 to 1.5. 

20200806, 13:46  #565  
Random Account
Aug 2009
U.S.A.
29·53 Posts 
Quote:
I ran a lot of P1's in the preGPU days despite not being able to make a direct connection between P1 bounds and TF start and stop bits. 

20200806, 13:54  #566 
"Oliver"
Sep 2017
Porta Westfalica, DE
307 Posts 
I guess in this case, your topic was "ignored" because no one had a good answer. Especially, when P1 are going to be reevaluated now.

20200806, 15:34  #567 
Random Account
Aug 2009
U.S.A.
3001_{8} Posts 
I understand what you mean by reevaluated, in this case. Their necessity, now that GPU's are becoming more powerful all the time. Recently, I have seen P1 tests which indicate the exponent was trialfactored to 78 bits. 80, and beyond, TF's will not be all that far behind for those who can run them in what they feel is a practical period of time.

20200806, 15:42  #568 
"Oliver"
Sep 2017
Porta Westfalica, DE
307 Posts 

20200806, 16:17  #569 
"James Heinrich"
May 2004
exNorthern Ontario
3^{2}×337 Posts 
I don't think that's what kruoli meant, but rather the the calculus of optimal P1 bounds (and TF limits) will change with the new PRP+Cert worktype that should, eventually replace LL+DC (or equivalent) with a single test and small verification, thereby eliminating (approximately) half the effort, therefore the effort applied to TF and P1 should be approximately half what it currently is (in broad terms).

20200806, 23:47  #570  
Random Account
Aug 2009
U.S.A.
29·53 Posts 
Quote:
Quote:


20200807, 00:37  #571 
6809 > 6502
"""""""""""""""""""
Aug 2003
101×103 Posts
2^{2}×5×431 Posts 
Remember that each bit level of TF is twice the effort of the one before it. So, 1 bit level less. And that leaves more factors in the low end of the P1 search to be found.
We live in interesting times. 
20200807, 14:55  #572 
Random Account
Aug 2009
U.S.A.
3001_{8} Posts 

Thread Tools  
Similar Threads  
Thread  Thread Starter  Forum  Replies  Last Post 
Prime95 version 29.2  Prime95  Software  71  20170916 16:55 
Prime95 version 29.1  Prime95  Software  95  20170822 22:46 
Prime95 version 26.5  Prime95  Software  175  20110404 22:35 
Prime95 version 25.9  Prime95  Software  143  20100105 22:53 
Prime95 version 25.8  Prime95  Software  159  20090921 16:30 