Perpetual benchmark thread...

[petrw1]

The Core i7 or at least, any current motherboard for the Core i7 as of November 13, 2008 (two ASUS boards, the intel DX58SO, gigabyte board -- data available from manufacturer website) does not support ECC memory.[citation needed] Some experts, such as Daniel Bernstein[10], recommend that systems without ECC support not be used for scientific computing, and not in general unless the user does not mind errors in critical data.

[mdettweiler]

Originally Posted by petrw1

The Core i7 or at least, any current motherboard for the Core i7 as of November 13, 2008 (two ASUS boards, the intel DX58SO, gigabyte board -- data available from manufacturer website) does not support ECC memory.[citation needed] Some experts, such as Daniel Bernstein[10], recommend that systems without ECC support not be used for scientific computing, and not in general unless the user does not mind errors in critical data.

Pardon my stupidity but...what the heck is ECC?

(Yeah, I know, I should probably Google this, but I'll probably get a better and clearer answer here. )

[Phantomas]

Pardon my stupidity but...what the heck is ECC?

best answer I found so far http://en.wikipedia.org/wiki/Dynamic...ror_correction

[Meikel]

More numbers!? OK... scalability of the Core i7.

Of course, you are right, stars10250, normally I also would run multiple different LL tests and not one multi-threaded one. So, how does that scale?

I don't use the benchmark for that, but use my regular work-units; double-checking in the 23000000-range currently (FFT-Length 1280K). I let it run for a minute or so and quote the best per-iteration time here. Then I'll start further tests on the other cores and watch, if and how the times for the first test increase.

This setting in local.txt guarantees, that I only use the real CPUs and not the hyperthreaded ones (HT was activated in BIOS):
[Worker #1]
Affinity=0
ThreadsPerTest=1

[Worker #2]
Affinity=2
ThreadsPerTest=1

[Worker #3]
Affinity=4
ThreadsPerTest=1

[Worker #4]
Affinity=6
ThreadsPerTest=1

And these are the results:

1 Test : Per iteration time 23.057ms
2 Tests: Per iteration time 23.714ms
3 Tests: Per iteration time 23.988ms
4 Tests: Per iteration time 24.240ms

I think, thats not too shabby. But now I am interested in Hyperthreading. What if I start 8 different tests? Of course, the times are expected to go up significantly. But I would hope for something a bit LESS than factor 2, as HT is supposed to improve utilization of every core. So, I add this to local.txt:
[Worker #5]
Affinity=1
ThreadsPerTest=1

[Worker #6]
Affinity=3
ThreadsPerTest=1

[Worker #7]
Affinity=5
ThreadsPerTest=1

[Worker #8]
Affinity=7
ThreadsPerTest=1

and let it go.

8 Tests: Per iteration time 48.920ms

So, thats 200% workload on each core and the result is about 200% time on each core. No gain from HT. But on the other hand: Not much of a loss, which is interesting, because with 2 threads per core, each one only has 256K of L2-cache!

Maybe it would be optimal to do 4 different tests and have each test run on two threads on the same core? That should probably avoid the L2-cache problem, because the two threads on the same FFT would also use the same L2-data, right? Unfortunately, the second worker threads can't be assigned to CPU numbers, and I can observe, that they tend to jump from core to core.

Can anyone tell me, how to assign the affinity of the helper thread?

[mdettweiler]

Originally Posted by Phantomas

best answer I found so far http://en.wikipedia.org/wiki/Dynamic...ror_correction

Ah, thanks for the link! It was quite enlightening.

Based on what this article said, though--namely, that most modern computers are *not* equipped with ECC memory--it would seem that Core i7 motherboards' not supporting ECC memory is not at all uncommon, and thus this would not render them any less suitable for GIMPS use.

[stars10250]

Thanks for the new i7 numbers Meikel. Your results seem to indicate true quad performance, in which all 4 cores have nearly identical throughput.

I'd like to build a new number cruncher (~$1k) so I've been sketching out the performance tradeoffs of oc i7 vs oc penryn vs a non-oc q6600 and I came to a funny conclusion. The first two systems are close in price and performance, depending on the build and oc, and can achieve nearly twice the performance of a non-oc q6600 (with its memory bottleneck)...quite a feat. But the q6600 system parts are becoming so cheap (even before i7 release) that one could buy three systems for the price of either the oc i7 or oc penryn. And three q6600s will outperform either of the other two systems readily. Yes, three will consume more power than one (which is a non-negligible cost), and you have to deal with everything times three, but it surprised me. The oc'd i7 and penryn require expensive cpu's & motherboards, expensive DDR3 ram, a video card, quality power supplies, sata HD, etc. I'm leaving out a bunch of details here, but I hope you see my point. The only thing is, my girlfriend is going to kill me if I order three more computers!

[petrw1]

Originally Posted by stars10250

Yes, three will consume more power than one (which is a non-negligible cost), and you have to deal with everything times three, but it surprised me. The oc'd i7 and penryn require expensive cpu's & motherboards, expensive DDR3 ram, a video card, quality power supplies, sata HD, etc. I'm leaving out a bunch of details here, but I hope you see my point. The only thing is, my girlfriend is going to kill me if I order three more computers!

I have a similar situation here. My employer (a large Government corp.) has a PC refresh policy that gives everyone a BRAND NEW, CURRENT PC every 3 or 4 years ... as for the old PC's (not that old): some are given to schools and the rest are offered for sale to the public (employees included) cheap. For example they are now removing and replacing 3 Ghz single core PCs and selling them (monitor and mouse included) for about $200.

I just ordered a new home PC (Q9550) but first toungue-in-cheek suggested to my wife that I could get 10 slightly used PCs from work for the same price. She would just have to clear out the storage room and install a second air conditioner and separate power box there first. I too was unsuccessful in convincing her. Realistically, though I don't really want a whole storage room full either.

[fivemack]

Originally Posted by petrw1

The Core i7 or at least, any current motherboard for the Core i7 as of November 13, 2008 (two ASUS boards, the intel DX58SO, gigabyte board -- data available from manufacturer website) does not support ECC memory.[citation needed] Some experts, such as Daniel Bernstein[10], recommend that systems without ECC support not be used for scientific computing, and not in general unless the user does not mind errors in critical data.

I am not convinced by this; for example, I have just run a very error-sensitive calculation for 850 hours on a quad-core Phenom without ECC, and it completed without trouble.

[fivemack]

Originally Posted by stars10250

Thanks for the new i7 numbers Meikel. Your results seem to indicate true quad performance, in which all 4 cores have nearly identical throughput.

I'd like to build a new number cruncher (~$1k) so I've been sketching out the performance tradeoffs of oc i7 vs oc penryn vs a non-oc q6600 and I came to a funny conclusion. The first two systems are close in price and performance, depending on the build and oc, and can achieve nearly twice the performance of a non-oc q6600 (with its memory bottleneck)...quite a feat. But the q6600 system parts are becoming so cheap (even before i7 release) that one could buy three systems for the price of either the oc i7 or oc penryn.

Have you also factored in that at least two of the three Q6600 don't need a hard disc or graphics card? I think it'll be a while before the best price-performance doesn't come from the cheapest available quad-core - though I admit that I have three at the moment, and getting another three would make the study Very Full and Very Loud.

[db597]

Originally Posted by stars10250

... oc i7 vs oc penryn vs a non-oc q6600

Why wouldn't you oc the q6600 as well? In fact, it's acknowledged to be an even better overclocker than the penryns. Most are Prime stable at 3-3.2Ghz with only a minimal bump in voltage. If you consider there's easily another 30% performance for free in there, the value proposition of the q6600 is even better.

[stars10250]

fivemack, I've never booted xp from a network so I'm not sure how to run them diskless. I looked on google, and it didn't seem straightforward. I have lots of old IDE drives I could use, though I do still like the diskless idea as it would be tidy and draw less power. If you have any good instructions, please post a link. I'm reasonably computer savvy, I've just never done it. I'd be using xp (my xp pro disk broke!). I ran ubuntu for a bit on one of my machines, but I disliked the feedback I got with prime95 so I reverted to xp. The motherboard that I was considering has built in video, so that's not an issue.

db597, I'm reluctant to oc because I've never done it and generally one needs a more expensive motherboard to do so. With the more expensive motherboards, they tend to use better ram (I was considering cheap DDR2 PC6400), and then often a video card. I suppose there are tons of instructions for the oc'd q6600 by now, so I guess I should try it. In my price estimate, I was considering an intel motherboard that costs around $85 which people have said doesn't oc. If you know an inexpensive one to use instead, let me know. I'll start researching this, as I've also heard good things about its oc ability.

On convincing my girlfriend, she said it was ok as long as I got rid of an equivalent number of older computers (which I was going to do anyway!). Power/cooling is an issue, but not right now in the cold months :)