Dream PC

[plandon]

OK, the scenario is;

You have won the lottery, or
You have received 3 months tax rebate, or
Your college has a 3% above inflation hardware budget, or
The NSA awards you $150,000 for a 100M prime, or
Your boss is away on indefinate mental health leave and you must spend the departments hardware budget before the end of the month, or
You claim your expenses from Westminster (UK joke)

What is the best PC for chasing Mersennes?

not the best PC for 600fps DeathMegaSplatter_TM or the best PC with enough memory to write a 2 paragraph letter to the bank manager using Microsoft software.

I know the answer changes every 2 weeks or the 17th June (whichever is smaller), so it needs to be frequently asked.

What is the best PC available for
A) Mersenne hunting
B) Racing to verify a new prime
C) other) GNFS?

Money is not infinite, but we are are not looking for the most productive per dollar either; more like asking the question
What is the current leading edge?

[mdettweiler]

Well...it depends on a number of factors. If you're limiting yourself to only one machine, then you'd want to get as big a machine as you can afford, of course. In this case, you'd probably want to go with big server components that can stick multiple Core i7's on one mobo. In fact, if you wait until 8-core Nehalems come out, then that would be even better.

If you're just trying to get the most bang-for-the-buck, however, then I'd recommend getting machines that are as cheap as you can get while still not sacrificing performance too much. For example, I put together some parts lists a while back for a friend of mine who was able to buy 4 Q6600 boxes for about $400 apiece. At the time, the top-of-the-line desktop CPU was about $500-$600, and that was just for the CPU. Obviously, getting multiple slightly-less-than-top-of-the-line machines was the best way to get the most raw computing power for the buck.

Of course, the biggest factor this depends on is exactly what you'll be using the machine for. Your three scenarios have rather different "ideal bang for the buck" combinations. In my friend's example, he's using the machines primarily to search for Riesel primes (k*2^n-1). The priorities used for picking out a machine for this are essentially identical to those one would use for picking out a machine for general GIMPS searching.

So, essentially, here's the factors you'd want to look for to get ideal bang for the buck on each of your three scenarios:

A) Mersenne hunting. As stated above, for this it's better to get more slightly subpar machines than to get one really great machine. Having 4 machines at 2.0 GHz is much better than one machine at 3.0 Ghz. Right now, the CPU's you'd want to get for this would probably be borderline between C2Q's and i7's.

B) Racing to verify a new prime. For this, you'd want to get one machine with as much raw CPU power as you can pack into it. The 8-core Nehalems would be perfect for this when they come out.

C) GNFS/etc. factoring. You want decent CPU power, but most importantly, LOADS of RAM. For GNFS/SNFS postprocessing of large jobs, it is essential to have a lot of memory. Of course, if the machine is mainly going to be used for GNFS/SNFS sieving (as opposed to postprocessing) then the RAM isn't quite as much of a priority and you'd want to focus more on the CPU. Nonetheless, though, even sieving takes a lot of RAM, so you'd still want to make sure you've got plenty of it. Also of note, a 64-bit system would be key, even more so than for the other priorities. NFS sieving goes twice as fast on 64-bit and you'll need access to the extra memory for both stages. Oh, and don't forget to get plenty of hard drive space: those relation files can be enormous.

Anyway, these are just approximate guidelines, but nonetheless the general idea holds true even as new technology is developed. Only the specifics change.

Max

[TimSorbet]

I have nothing more to add, except to ask the OP if this idea is limited to one PC or can be a farm for the best performance/cost ratio.
p.s. When looking at performance/cost ratios, remember electricity costs.

[mdettweiler]

Originally Posted by Mini-Geek

I have nothing more to add, except to ask the OP if this idea is limited to one PC or can be a farm for the best performance/cost ratio.
p.s. When looking at performance/cost ratios, remember electricity costs.

Actually, I did address the question of one PC vs. multiple PC's in my above post. Though, yes, you're right, electricity is definitely an important consideration. In fact it could be considered as part of the larger consideration of one PC vs. multiple; if electricity is not a big deal, then multiple machines are the way to go, and vice versa. One honking big machine will almost definitely use less power than four not-so-honking machines.

[plandon]

I am thinking of the dream systems, not the most bang for the buck.
The electricity bill is not a problem.

For example for System B:
So who makes the best 8x8 core EX mobos?

What is the most hardcore PC available for System A?
Alienware seem to have lagged recently, what about a top of the range White Shark.
Can anybody else overclock more than 4GHz?

[mdettweiler]

Originally Posted by plandon

I am thinking of the dream systems, not the most bang for the buck.
The electricity bill is not a problem.

For example for System B:
So who makes the best 8x8 core EX mobos?

What is the most hardcore PC available for System A?
Alienware seem to have lagged recently, what about a top of the range White Shark.
Can anybody else overclock more than 4GHz?

If you're talking just one big powerful system, then the ideal requirements for both A and B are essentially the same, and thus you'd want to go with what I described for B. The only difference is in how you'd actually utilize the machine once it's built; for ordinary Mersenne hunting, you'd get the best throughput by running one test per core for 16 concurrent tests (or, 8 tests split up over the hyperthreaded cores will be approximately similar performance-wise). For fast verification, though, you'd want to do 16 threads on one test.

[joblack]

The nehalem ex (or the new 6 core amd) is only interesting for the hunt for 100M numbers. For everything else you need no high multicore machine. If you have won in the lottery get 20 of the intel 5400 double core cpus ... costs around 150 Euro for a machine.

[joblack]

Of course with a big lottery win you can get a multicore nehalem ex (16 or more cores) to go for the 100M challange.

[lavalamp]

I'd buy the game DeathMegaSplatter.

Best regular desktop machine - overclocked i7 965
Best server machine - quad socket mobo populated with 4*Xeon W5580s **

The server would be EXTREMELY expensive, it might acually work out cheaper to buy four of the desktop machines instead. But if you did that I guess you couldn't run Prime95 with 32 threads (if it even supports that many). I wonder what the throughput would be like on that many threads, the returns might very well be thoroughly diminished by that point.

** Haven't actually seen a quad Nehalem board, they may only be made for dual Nehalem CPUs at the moment.

[henryzz]

http://www.mersenneforum.org/showpos...&postcount=103
look at this for scalablity
prime95 is so rubbish at it that glucas(i am pretty certain the slowest of the 3 major programs with 1 core) beats it with 4 cores

[ATH]

Originally Posted by joblack

Of course with a big lottery win you can get a multicore nehalem ex (16 or more cores) to go for the 100M challange.

But can it play Crysis 2 !?!?!??!?