Most efficent cores to use for 100million LL - Page 2

Batalov · 2014-06-26, 00:06

Quote:

Originally Posted by retina

In fact you may have already experienced this but the flipped bit(s) were in unused memory and thus had no affect.

Aw, the famous black cat in the pitch black room argument! Love it!

TheMawn · 2014-06-26, 00:14

I'm saying that there is always that unknown chance for a single bit error. Overclocking has nothing to do with this.

What I am trying to say is you can't ever guarantee that a mistake will not be made. You can only become reasonably confident, and that takes time.

Suppose you do twelve months of DC's to stability test your system, and you got zero errors. All you can really say is your odds of failure for one month are less than one in twelve, if you consider each month to be statistically significant. In other words, your odds of succeeding for one month are greater than 91.7%.

Now, [Greater than 0.917]¹² = [Greater than 35.2%]. These are your odds of getting twelve consecutive successes.

If I tell you that your odds of having a successful LL are better than a third, are you really inspired with confidence?

"Now wait, I might have gone twelve months without an error, but did I not also go for 30 million seconds? My odds of failing any particular second are minuscule, so doesn't the interval length matter?"

Of course it does. However, this boils down to something very familiar. Some people reading this might already know where I'm going.

(1 - 1/30,000,000)^30,000,000 = 36.79%. Your odds of not failing an iteration for 30,000,000 seconds are still only better than a third. Note that they could be anywhere above a third, including 100.000%. But that's all we can say for sure.

You can keep making your results better by taking smaller intervals, but your odds of success are simply tending to e^-1.

Now, if you go for 24 months without an error, your odds of success per month are > 95.8%. Raise that to 12 (you're only doing a one-year test) and you get > 60%. Want to try seconds? You should get > 60.6%.

If you want a formula, your odds of a successful year-long test after x years without a failure are only greater than e^-1/x.

60.6% after two years.
81.8% after five years.
99.0% after 100 years.

EDIT: I started writing this and was interrupted, and there were some replies in the mean time. Retina of course makes a good point: The larger tests have more chances of failure per length of time by themselves. So really, your chances of succeeding a one-year test after two years of testing are less than greater than 60.6%

kladner · 2014-06-26, 01:23

Just to backtrack-

Quote:

Every single DC I have done with this CPU has come up with a matching residue

This is shear luck. You have DC'd only correct first time tests. (This is much easier to accept than the idea that you and the First Tester(s) independently arrived at the same wrong residue.) Not matching the residue does not prove one way or the other about the correctness of one's machine's performance.

potonono · 2014-06-26, 03:09

I protect my calculations from cosmic rays with a thick layer of dust over all the internal PC components.

retina · 2014-06-26, 03:14

Quote:

Originally Posted by potonono

I protect my calculations from cosmic rays with a thick layer of dust over all the internal PC components.

Except that your layer of dust has impurities and emits alpha particles thus nullifying the effect you seek to obtain.

TheMawn · 2014-06-26, 03:21

What I am trying laboriously to say, though certain individuals persist in trying to misunderstand me, is that as a personal example, despite my CPU having as clean a record as possible over a year, I would not trust it with a 100 million digit LL.

kracker · 2014-06-26, 15:17

Quote:

Originally Posted by TheMawn

What I am trying laboriously to say, though certain individuals persist in trying to misunderstand me, is that as a personal example, despite my CPU having as clean a record as possible over a year, I would not trust it with a 100 million digit LL.

I think you're misunderstanding us. Who says/why do we have to test our CPU's for over a year? Simple, stay at stock. Do you really think CPU's are that error prone? A 332M test takes around 120 days on a Haswell with 4 threads.
Maybe I am missing something.

retina · 2014-06-26, 15:31

Quote:

Originally Posted by kracker

Maybe I am missing something.

ECC.

kladner · 2014-06-27, 01:28

Quote:

Originally Posted by retina

ECC.

The longer the test, the more you need it.

2014-06-26, 00:14	#13
TheMawn May 2013 East. Always East. 3277₈ Posts	I'm saying that there is always that unknown chance for a single bit error. Overclocking has nothing to do with this. What I am trying to say is you can't ever guarantee that a mistake will not be made. You can only become reasonably confident, and that takes time. Suppose you do twelve months of DC's to stability test your system, and you got zero errors. All you can really say is your odds of failure for one month are less than one in twelve, if you consider each month to be statistically significant. In other words, your odds of succeeding for one month are greater than 91.7%. Now, [Greater than 0.917]¹² = [Greater than 35.2%]. These are your odds of getting twelve consecutive successes. If I tell you that your odds of having a successful LL are better than a third, are you really inspired with confidence? "Now wait, I might have gone twelve months without an error, but did I not also go for 30 million seconds? My odds of failing any particular second are minuscule, so doesn't the interval length matter?" Of course it does. However, this boils down to something very familiar. Some people reading this might already know where I'm going. (1 - 1/30,000,000)^30,000,000 = 36.79%. Your odds of not failing an iteration for 30,000,000 seconds are still only better than a third. Note that they could be anywhere above a third, including 100.000%. But that's all we can say for sure. You can keep making your results better by taking smaller intervals, but your odds of success are simply tending to e^-1. Now, if you go for 24 months without an error, your odds of success per month are > 95.8%. Raise that to 12 (you're only doing a one-year test) and you get > 60%. Want to try seconds? You should get > 60.6%. If you want a formula, your odds of a successful year-long test after x years without a failure are only greater than e^-1/x. 60.6% after two years. 81.8% after five years. 99.0% after 100 years. EDIT: I started writing this and was interrupted, and there were some replies in the mean time. Retina of course makes a good point: The larger tests have more chances of failure per length of time by themselves. So really, your chances of succeeding a one-year test after two years of testing are less than greater than 60.6% Last fiddled with by TheMawn on 2014-06-26 at 00:18

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
32 cores limitation	gabrieltt	Software	12	2010-07-15 10:26
CPU cores	Unregistered	Information & Answers	7	2009-11-02 08:27
Running on 4 Cores	Unregistered	Information & Answers	9	2008-09-25 00:53
6 Intel Cores	petrw1	Hardware	3	2008-09-16 16:33
A program that uses all the CPU-cores	Primix	Hardware	7	2008-09-06 21:09

2014-06-26, 03:09	#15
potonono Jun 2005 USA, IL 193 Posts	I protect my calculations from cosmic rays with a thick layer of dust over all the internal PC components.

2014-06-26, 03:21	#17
TheMawn May 2013 East. Always East. 11×157 Posts	What I am trying laboriously to say, though certain individuals persist in trying to misunderstand me, is that as a personal example, despite my CPU having as clean a record as possible over a year, I would not trust it with a 100 million digit LL.