Best Unit for the Job?

[cheesehead]

Originally Posted by joblack

With Murphy's Law you'll do

all that and still not find a Mersenne prime.

[joblack]

Originally Posted by Flatlander

I think Murphy's Law does apply here. It states that the box will be blown by lightning at 99% of a 100M digit test. (Could actually be more likely than the result being prime.)

In that exponent range I would at least make daily backups (if not building a RAID system). Losing your data at 99% would be really annoying.

[Primeinator]

Is a super-fast hard drive really that necessary for prime hunting though? DDR3 makes sense...faster and better RAM. How good was the system I posted a link to?

[joblack]

Originally Posted by cheesehead

all that and still not find a Mersenne prime.

If you would like to use Prime95 in a realistic manner you wouldn't use it anyway. Don't be that pessimistic :D ...

[joblack]

Originally Posted by Primeinator

Is a super-fast hard drive really that necessary for prime hunting though? DDR3 makes sense...faster and better RAM. How good was the system I posted a link to?

No it isn't ... a cheaper ASRock board is enough. You also don't need more than 4 Gigs RAM but a 64 Bit OS is better (gives you a 10 percent speed advantage because of the additional SSE registers). Think about creating a RAID1 (included in the board - just use two harddisks) and an additional USB Stick for backup purposes. For overclocking don't go over 3,7 Ghz (for the 920) because over that the Watt/Performance ratio decreases (check tomshardware.com for more overclocking information) ...

[CRGreathouse]

Originally Posted by joblack

If you would like to use Prime95 in a realistic manner you wouldn't use it anyway. Don't be that pessimistic :D ...

I think that was just poking fun at your (mistaken) invocation of Murphy's Law.

[lavalamp]

Originally Posted by joblack

You mean the first 100M digits prime? Perhaps its only 2 years (it's hard to estimate the time of the Core i7 if you don't have one) - the chance to find a 10M prime numbers isn't that much higher.

No, read what he said again.

If you want the best desktop PC, get an i7 975 with 12 GB DDR3 2000 MHz RAM and overclock it.

If you want a great desktop PC get an i7 920 (make sure it's one with the D0 stepping) with 3 GB DDR3 1333 MHz RAM, or better yet 1600 MHz RAM. Overclocking would net you more performance, but you'd need an aftermarket cooler and you SHOULD NOT attempt this without research, you can easily burn out the motherboard or CPU if you do not know what you are doing. I'd say most people who overclock don't kill their equipment more out of luck than skill.

If you want a cheap but still good desktop PC, get a low end 45nm Core2Quad with 2 GB DDR2 1066 MHz RAM, there should be some overclocking headroom even with the default cooler and they are easier to OC than the i7s.

If you want the best performance per watt, I'd say go with the above Core2Quad system, and downclock it.

The Phenom and Phenom II series CPUs from AMD will offer decent performance too, but the highest end Phenom II CPU still only performs about on par with the lowest i7.

[CRGreathouse]

Originally Posted by lavalamp

The Phenom and Phenom II series CPUs from AMD will offer decent performance too, but the highest end Phenom II CPU still only performs about on par with the lowest i7.

The Phenom II is a great chip -- I'm running one right now. But the original Phenom isn't worth buying, especially with how cheap a Phenom II is:

• $190 for 3.0 GHz
• $140 for a 2.6 GHz

And don't forget that motherboards for the Phenom II are cheaper than mobos for the i7 (save $80 to $150, assuming you use an AM2+ rather than AM3) and that you can use DDR2 instead of DDR3 (save $100).

Overall a Phenom II system can be $320 to $350+ cheaper than a similarly-clocked i7-based system. (Of course Prime95 is better optimized for the i7, so if that's what you're doing it's better to compare to a Phenom II running at a higher speed.)

But a Phenom system will save only about $50 over the Phenom II, and its actual performance is much lower (cache issues, Cool n' Quiet problems, etc.).

[rudi_m]

Also, what sites are the best for custom building such a machine?

Of course I would not spent money just to run mprime faster.

But for example if you are interested in some kind of hosting webservices anyway, then I would rent such fast machine.
In Germany its very cheap, here for example:
http://www.hetzner.de/de/hosting/pro...produktmatrix/
Intel® Core™ i7-920
8GB RAM
49 EUR /month
IMO very cheap - just running such thing 24/7 at home would cost you ~20 EUR electricity per month plus buying the hardware.
And there you have additionally a 100 Mbit connection, traffic unlimited to use it for a lot of other things.
Also if it produces a Bad LL - they will give you a new machine within some hours.
On next cpu generation you can just switch to the next special offer.

[Primeinator]

So a quad Phemon II at 3.0 Ghz would be a good choice?

How good is this system? http://www.tigerdirect.com/applicati...6473&CatId=333

[R.D. Silverman]

Originally Posted by Primeinator

Changing my mind from a previous thread, I am just going to save up and buy a more powerful crunching unit instead of just a new CPU. So...my question is this. Without being prohibitively expensive, and just looking for a computer that will be used mainly for number-crunching, what is the best combination of CPU, RAM, Motherboard, etc for performing large LL tests 45M plus and eventually going for the hundred million digit challenge? Also, are there any CPUs coming out in the next year or two that are going to be breakthroughs so to speak on LL speed, etc? Thanks.

Also, what sites are the best for custom building such a machine?

The question, as asked, is poorly conceived. The following
will be an intro to economic decision analysis. The question
you ask can be modelled as a linear program.

(1) What are the constraints? I assume you have a constraint
(i.e. upper bound) on cost. Do you consider just the purchase
price or do you include operating costs? Do you have a constraint
on your time? Managing multiple systems as opposed to a single
system will require more time. Multiple systems will have greater
operating costs. It might be cheaper to (say) buy 4 lesser systems
(which run at say 1/3 to 1/2 the speed) or a single very fast
system. The former will have greater throughput, the latter
lower latency.

(2) Maximizing "bang for the buck" is meaningless. You can
always get more performance by spending more. OTOH, you can
probably find a 10-year old used system for $5.00, which will
be 50 times slower than a new system. But at only $5.00 you
would be getting greater "bang for the buck" (i.e. price
performance)

(3) Are you looking to maximize price/performance? (with say a
lower bound on performance?)
Are you looking to maximize throughput?
Are you looking to minimize latency (time to test a single
number)?

These are not the same question.


Specify your variables. Specify your constraints. Specify your objective.

Variables will be a measure of perfomance for each system and the cost of
each system. Another variable might be the electrical costs.
Constraints will be a lower bound on performance, perhaps a bound on
price/performance and an upper bound on costs. The objective is up
to you.

Solve.
Then do a sensitivity analysis. Compute your shadow prices. Might
you be willing to relax your upper bound on cost by e.g. 10%, if
it means a 20% improvement? etc. etc.


FINALLY: You can always get a better system for the same price by
waiting e.g. a year. This should never be a consideration unless your
objective is to maximize throughput over a *fixed time period* starting
today. i.e. if you want to test the most numbers over the next 5 years
it might pay to wait.