Prime95 and graphics cards - Page 2

Dubslow · 2011-12-06, 16:09

Indeed. Maybe 72 bits for 55M+ -- but not 50M or lower. Certainly not the lower ranges of GPU272. (At least IMO -- OTOH, our work also depends on how fast the LL wavefront is moving, and depending on how slow that is...)

chalsall · 2011-12-06, 16:28

Quote:

Originally Posted by Dubslow

Indeed. Maybe 72 bits for 55M+ -- but not 50M or lower. Certainly not the lower ranges of GPU272. (At least IMO -- OTOH, our work also depends on how fast the LL wavefront is moving, and depending on how slow that is...)

I'm working on a report which should help us answer this question based on empirical data rather than gut feel. Basically, average cost per factor found for each bit level and 1M range.

This will help us determine when we should start going to 73....

diamonddave · 2011-12-06, 16:32

Quote:

Originally Posted by chalsall

I'm working on a report which should help us answer this question based on empirical data rather than gut feel. Basically, average cost per factor found for each bit level and 1M range.

This will help us determine when we should start going to 73....

or 70 for DC

Dubslow · 2011-12-06, 16:36

Scary.

OP, if you're still reading this, what do you think of the answer?

davieddy · 2011-12-06, 22:42

Quote:

Originally Posted by Dubslow

Scary.

OP, if you're still reading this, what do you think of the answer?

I'm afraid to say that I suspect the OP is running as fast as
(s)he can from us nutters

David

LaurV · 2011-12-07, 03:31

Quote:

Originally Posted by diamonddave

or 70 for DC

I already proved that even 69 is too much for DC. It takes more days to find a factor that it would take to CudaLucas to DC the exponent. Especially when all the available DC exponents HAD A LOT of P-1 done on them already, the missed factors are scarce. See my stats on the project: I only started to DC few days ago, I already cleared 11 exponents by DC-LL them (19, including some for which I found factors in my first days of TF-ing) with two GPU's (only one residue mismatched) and in another week I will clear another 20 or so, overtaking the guys with over 30 factors found. Of course, I will not get the billion GHz-days credit that mfaktc give, but the clearing is certain, it come with "two point something exponents cleared per day" and not with "maybe I will find a factor in the next two days, maybe not".

The goal is to clear the exponents as fast as possible.

DC@DCfront
TF@LLfront
P-1 on old CPU's
LL@LLfront on new CPU and/or GPU
No TF@DCfront

That is the key.

Dubslow · 2011-12-07, 06:25

Indeed, I was saying as much in another thread, but had no numbers to back it up :)

chalsall · 2011-12-07, 14:26

Quote:

Originally Posted by chalsall

I'm working on a report which should help us answer this question based on empirical data rather than gut feel. Basically, average cost per factor found for each bit level and 1M range.

The report is now available. Cost per Factor Found broken down by "bit depth" and 1M ranges.

Hopefully this will help us decide more authoritatively how deep we should go for each range.

diamonddave · 2011-12-07, 14:55

Quote:

Originally Posted by chalsall

The report is now available. Cost per Factor Found broken down by "bit depth" and 1M ranges.

Hopefully this will help us decide more authoritatively how deep we should go for each range.

I'm really surprised by those results. We are finding way more factor then expected, excepted for the 69 bit level in DC.

Also the Factor found reported here doesn't agree with other report on the site.

Cross posted in GPU to 72
LL:

Code:

Bit     Exp.    Factor Expected
69	1640	61	24
70	9949	250	142
71	6020	192	85
72	4244	112	59
73	78	3	1

DC:

Code:

Bit     Exp.    Factor  Expected
68	1477	32	22
69	5291	50	77
70	83	4	1

bcp19 · 2011-12-07, 19:44

Quote:

Originally Posted by LaurV

I already proved that even 69 is too much for DC. It takes more days to find a factor that it would take to CudaLucas to DC the exponent.

You may have "proven" it in regards to your GPU, but it is not true for all of them. My GPU's can test 172/176/189 exponents in the time it would take each to run a single DC. With 50 factors cleared for 5291 exponents, or 1 every 106 exponents, I am clearing exponents almost 70% faster TFing 68 to 69 than CUDALucas would.

bcp19 · 2011-12-07, 21:17

I also just tested a GT 520, which can test about 21 TF a day but would take just under 17 days to perform a DC, which is roughly 350 TF per DC. Looking at where a normal 560Ti falls on http://mersenne-aries.sili.net/mfaktc.php I would hazard a guess that only 2-4 GPU's are in the same "proven" category as yours.

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
New Linux rootkit leverages graphics cards for stealth.	swl551	Lounge	0	2015-05-08 14:06
Nvidia's next-generation graphics cards	ixfd64	GPU Computing	22	2014-11-15 04:25
Prime95 On Graphics Card	louis_net	Hardware	10	2004-11-26 22:05
how do graphics cards work so fast?	ixfd64	Hardware	1	2004-06-02 03:01
Chance to use modern Graphics Cards as..	Marco	Hardware	28	2003-11-02 23:21

2011-12-06, 16:09	#12
Dubslow Basketry That Evening! "Bunslow the Bold" Jun 2011 40<A<43 -89<O<-88 3×29×83 Posts	Indeed. Maybe 72 bits for 55M+ -- but not 50M or lower. Certainly not the lower ranges of GPU272. (At least IMO -- OTOH, our work also depends on how fast the LL wavefront is moving, and depending on how slow that is...)

2011-12-06, 16:36	#15
Dubslow Basketry That Evening! "Bunslow the Bold" Jun 2011 40<A<43 -89<O<-88 3·29·83 Posts	Scary. OP, if you're still reading this, what do you think of the answer?

2011-12-07, 06:25	#18
Dubslow Basketry That Evening! "Bunslow the Bold" Jun 2011 40<A<43 -89<O<-88 3×29×83 Posts	Indeed, I was saying as much in another thread, but had no numbers to back it up :)

2011-12-07, 21:17	#22
bcp19 Oct 2011 679₁₀ Posts	I also just tested a GT 520, which can test about 21 TF a day but would take just under 17 days to perform a DC, which is roughly 350 TF per DC. Looking at where a normal 560Ti falls on http://mersenne-aries.sili.net/mfaktc.php I would hazard a guess that only 2-4 GPU's are in the same "proven" category as yours.