Tweaking RAM & CPU - Page 4

lavalamp · 2011-01-11, 04:36

Quote:

Originally Posted by lorgix

Does OCing the PCI-E make any sense btw?

Absolutely none at all, and it's strongly recommended that you don't.

lorgix · 2011-01-11, 07:40

Quote:

Originally Posted by Mr. P-1

It's about finding as small a set of numbers as possible within B1,B2 that covers all the primes. The primorials have the property that the ratio of relative primes phi(p#) to p# is less than for any smaller number. Small multiples of p# (smaller than the next primorial) have the same ratio. So I can't see where the saving comes from in using them.

When you put it that way... Neither can I.

lorgix · 2011-01-11, 11:17

Does anyone have any advice on how to set Prime95's 'Torture Test' to be... well, as much torture as possible?

CPU data:

L1-D 32K * 2, 8-way
L1-I 32K * 2, 4-way

L2 256K * 2, 8-way

L3 4M, 16-way

Quote:

Originally Posted by lavalamp

Absolutely none at all, and it's strongly recommended that you don't.

Ok, I won't. Thanks for the advice.

Btw, would you know if it is at all possible to run DDR3 at a command rate of 1T?

lavalamp · 2011-01-11, 13:49

Try it and see, if the setting is there enable it, and if you're still stable then great. The difference is only slight though, 2-3% in terms of real world performance, be aware that artificial benchmarks that solely measure RAM speed and latency will greatly inflate the purported performance benefit.

As for the torture test, that depends what you want to torture. Max CPU stress is achieved with small FFTs, and max RAM stress is achieved with larger FFTs, and the blend test is somewhere in between. But Prime95 tells you all this in the window anyway.

If you're overclocking your CPU then you should hammer the CPU. If you're overclocking the RAM then similarly hit the RAM. Try not to do them both at the same time, which is sometimes hard, but handily you can, because you can overclock your CPU purely with the multiplier. The Intel burn test is better at stressing both the CPU and RAM simultaneously, which is why it's such a good stress tester, your whole system is put under maximum load.

lorgix · 2011-01-11, 18:11

I've tested these new settings thoroughly.

I'm quite confident it's stable at 21*160.

As assessed by Linpack, the FLOPS gain was 4.85%. Using 6GB.

cheesehead · 2011-01-14, 07:41

Quote:

Originally Posted by Mr. P-1

While this is not technically incorrect, (The ratio exists. It has a particular value at optimal B1,B2, which is different from its value at other B1,B2, and so could be considered the optimal ratio) it is quite misleading.

Oops, I once again substituted my mental model for the actual computation. Thanks for your correction.

James Heinrich · 2011-03-15, 12:29

Quote:

Originally Posted by Mr. P-1

I have never seen that number of relative primes before. 432 = 9 * 48 implying a blocksize of 9 * 210. How many relative primes per pass were you getting?
I have seen 192 relative primes before, but not often. IIRC, that happened when memory was relatively tight.

Do I correctly deduce from this thread that the number of relative primes is always a multiple of 48? I've seen mentioned in this thread 192, 432, 480, 2400; I have seen 960 before and I just noticed this running right now:

Quote:

Available memory is 9939MB.
Using 7724MB of memory. Processing 202 relative primes (86 of 288 already processed).
M80101757 stage 2 is 45.226701% complete.

I still don't grasp the relationship to number of relative primes to efficiency; when is it better for Prime95 to deviate (either up or down) from the standard 480?

James Heinrich · 2011-03-15, 15:41

Strangely, after months of seeing nothing but 480, I just saw another new one within a few hours of the above (different computer though):

Quote:

M334000013 stage 1 complete. 669922 transforms. Time: 57705.731 sec.
Stage 1 GCD complete. Time: 794.490 sec.
Using 6466MB of memory. Processing 39 relative primes (0 of 1440 already processed).

Mr. P-1 · 2011-03-19, 21:56

Quote:

Originally Posted by James Heinrich

Do I correctly deduce from this thread that the number of relative primes is always a multiple of 48?

Actually it's always a multiple of 8. If more than 48, always a multiple of 48. If more than 480, always a multiple of 480.

The magic numbers, 8, 48, and 480 are the values of Euler's totient function at 5#, 7# and 11# respectively. The next value: phi(13#) is 5760, but Prime95 doesn't use this.

Quote:

I've seen mentioned in this thread 192, 432, 480, 2400; I have seen 960 before and I just noticed this running right now:I still don't grasp the relationship to number of relative primes to efficiency; when is it better for Prime95 to deviate (either up or down) from the standard 480?

I have absolutely no idea why it should ever choose any value other than 8, 48, and 480.

Mr. P-1 · 2011-03-19, 22:01

Quote:

Originally Posted by James Heinrich

Using 6466MB of memory. Processing 39 relative primes (0 of 1440 already processed).

Based upon my understanding of the algorithm, that's an insane choice. You're paying a penalty for having 37 passes instead of 13 to no benefit whatsoever, as far as I can see.

James Heinrich · 2011-03-19, 22:27

Quote:

Originally Posted by Mr. P-1

Quote:

Using 6466MB of memory. Processing 39 relative primes (0 of 1440 already processed).

Based upon my understanding of the algorithm, that's an insane choice. You're paying a penalty for having 37 passes instead of 13 to no benefit whatsoever, as far as I can see.

I moved stage 2 off that machine to my other machine which has 10GB allocated, so now it's running 63 relative primes at once (instead of 39).

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Large FFT tweaking	Zerowalker	Information & Answers	8	2013-04-19 15:01
Tweaking polynomial search for C197	fivemack	Msieve	38	2011-07-08 08:12
Tweaking and compiling the Kleinjung siever	Batalov	Factoring	57	2010-11-30 18:03

2011-01-11, 13:49	#37
lavalamp Oct 2007 Manchester, UK 23·59 Posts	Try it and see, if the setting is there enable it, and if you're still stable then great. The difference is only slight though, 2-3% in terms of real world performance, be aware that artificial benchmarks that solely measure RAM speed and latency will greatly inflate the purported performance benefit. As for the torture test, that depends what you want to torture. Max CPU stress is achieved with small FFTs, and max RAM stress is achieved with larger FFTs, and the blend test is somewhere in between. But Prime95 tells you all this in the window anyway. If you're overclocking your CPU then you should hammer the CPU. If you're overclocking the RAM then similarly hit the RAM. Try not to do them both at the same time, which is sometimes hard, but handily you can, because you can overclock your CPU purely with the multiplier. The Intel burn test is better at stressing both the CPU and RAM simultaneously, which is why it's such a good stress tester, your whole system is put under maximum load.

2011-01-11, 18:11	#38
lorgix Sep 2010 Scandinavia 3×5×41 Posts	I've tested these new settings thoroughly. I'm quite confident it's stable at 21*160. As assessed by Linpack, the FLOPS gain was 4.85%. Using 6GB.