Cheapest CPU cycles?

[siegert81]

Assuming one has roughly $1,000 to blow on processing power for the sole purpose of number crunching, what's the most cost effective approach? What sort of computer(s) should one buy? What are the minimum components and parts?

Personally, I already have a computer, so I don't need a new monitor, speakers, keyboard, or mouse. I can borrow those parts from my current computer just to initialize the programs. I'd prefer to have a bare-bones system that just seeks prime numbers. How can I maximize my CPU cycles per dollar? Also, are monthly electricity costs significant?

[xilman]

Originally Posted by siegert81

Assuming one has roughly $1,000 to blow on processing power for the sole purpose of number crunching, what's the most cost effective approach? What sort of computer(s) should one buy? What are the minimum components and parts?

Personally, I already have a computer, so I don't need a new monitor, speakers, keyboard, or mouse. I can borrow those parts from my current computer just to initialize the programs. I'd prefer to have a bare-bones system that just seeks prime numbers. How can I maximize my CPU cycles per dollar? Also, are monthly electricity costs significant?

I have a number of questions, the answer to which will have a bearing on your decision.

1. What sort of number crunching? Something which needs oodles of RAM, such as the linear algebra phase of the NFS will be different from something which fits into L2 cache for much of its existence.
2. What do you mean by "seeks prime numbers"? What kind of prime numbers and what kind of seeking? Again, some methods benefit from having large amounts of memory available, others do not.
3. What constraints on electricity costs do you have? Where I live, and especially right now, I need to keep my house warm. Any power used by computational fan heaters does not need to be provided by alternative heating elements.
4. Is your targeted application well-matched to a SIMD machine? If so, you may find it cost effective to spend roughly 30% of your budget on a mid-range CUDA-capable GPU. This advice is particular appropriate if you are prepared to do your own programming.

Paul

[nucleon]

GPGPU!

At the moment, you can't beat GTX460 for value for money both initial costs and ongoings.

The downside is they need a bit of end user time. The current code isn't exactly fire and forget. (i.e. you need to manually get work, and submit results manually)

Here in Australia I priced a 6-core AMD solution with 2x NV460GTX-1G for under $1050 (Note: this is local prices and currency, your local prices may vary)

That's without luxuries like a case. :)

-- Craig

[joblack]

The AMD 6-core CPUs have the most 'bang for the buck'. If you want to invest more money for a little bit more performance you might consider a Corei7 Quad Core processor.

[siegert81]

To answer your questions:

I'd like to run NewPGen & LLR in order to find proth primes or possibly help out in the Riesel problem (using BOINC/PrimeGrid). Ideally, I'd like to find large proth primes with low k values, which will give me a decent chance of discovering a Fermat factor.

I am not prepared to do my own programming. However, manually creating work and submitting primes to the top 5000 database is perfectly fine. If I were to get a GPU, I'd just need to know which program(s) to use and how to use them.

As for total cost, I see now that the $1,000 figure was misleading. I am interested in maximizing the ratio of CPU cycles to dollars spent. If spending $2000+ produces a higher ratio (more bang/$), then so be it.

In another thread I read that the GPUs are better for factoring. Can they perform LLR tests as well? In fact, I was just watching CNBC a minute ago. They were discussing the company NVidia, and they mentioned their GPUs being great for various research projects. Ha! Anyhow, I currently do not have a GPU for my dual-core 3.2 GHz desktop. Would this be a good place to start in order to get a lot of CPU cycles for my money?

[ET_]

Originally Posted by siegert81

To answer your questions:

I'd like to run NewPGen & LLR in order to find proth primes or possibly help out in the Riesel problem (using BOINC/PrimeGrid). Ideally, I'd like to find large proth primes with low k values, which will give me a decent chance of discovering a Fermat factor.

I am not prepared to do my own programming. However, manually creating work and submitting primes to the top 5000 database is perfectly fine. If I were to get a GPU, I'd just need to know which program(s) to use and how to use them.

As for total cost, I see now that the $1,000 figure was misleading. I am interested in maximizing the ratio of CPU cycles to dollars spent. If spending $2000+ produces a higher ratio (more bang/$), then so be it.

In another thread I read that the GPUs are better for factoring. Can they perform LLR tests as well? In fact, I was just watching CNBC a minute ago. They were discussing the company NVidia, and they mentioned their GPUs being great for various research projects. Ha! Anyhow, I currently do not have a GPU for my dual-core 3.2 GHz desktop. Would this be a good place to start in order to get a lot of CPU cycles for my money?

If you are searching for Riesel and/or Fermat factors, you can use ppsieve with a nVidia GPU to sieve out composites. I think there should be a Boinc-aware platform that uses nVidia GPUs.

Luigi

[mdettweiler]

Originally Posted by ET_

If you are searching for Riesel and/or Fermat factors, you can use ppsieve with a nVidia GPU to sieve out composites. I think there should be a Boinc-aware platform that uses nVidia GPUs.

Luigi

Indeed--PrimeGrid (the OP's project of choice) supports CUDA (nVidia) GPUs natively via its PPS Sieve subproject (which uses tpsieve to sieve Riesel and Proth numbers simultaneously). The process is automatic; just install the CUDA drivers and then set BOINC's options to "Use nVidia GPU->Yes".

@siegert81: first of all, there is no point in using NewPGen for sieving nowadays except for some obscure search types. For Proth and Riesel primes, your best bet is ppsieve/tpsieve or sr2sieve (depending on whether you want to sieve a large block of k's or a scattered handful, respectively). The former is covered quite fully in the form of PrimeGrid's PPS Sieve subproject.

To maximize your bang for the buck, what I would do is buy two AMD Phenom II X6 machines (1055T or 1090T) which together should total about $900-$1000. If you add an nVidia GTX 460 (make sure to get the version with 1 GB memory) to each, the total should come to about $1200-1400 or so. That, I'm estimating, would give you a reasonably optimal "bang for the buck" setup. (Alternatively, you could go with Intel Core i5 CPUs, which are faster per-core than the AMDs; but the AMDs' two additional cores more than make up for the difference. Note that the i5s overclock rather better as I understand it, so if you're planning to try that then it becomes more of a toss-up.)

Once you have the two machines, you'd have two fast GPUs to work on the PPS Sieve subproject, and 12 reasonably fast CPU cores to work on LLR subprojects (PPS LLR and TRP LLR as you indicated). That kind of setup will crank out quite a bit of work on both fronts.

Edit: to answer another of your questions, GPUs can not yet perform LLR tests. Hopefully they will at some point in the near future, but for now, you're best off sieving on GPUs, and LLRing on CPUs.

[xilman]

Originally Posted by mdettweiler

To maximize your bang for the buck, what I would do is buy two AMD Phenom II X6 machines (1055T or 1090T) which together should total about $900-$1000. If you add an nVidia GTX 460 (make sure to get the version with 1 GB memory) to each, the total should come to about $1200-1400 or so. That, I'm estimating, would give you a reasonably optimal "bang for the buck" setup. (Alternatively, you could go with Intel Core i5 CPUs, which are faster per-core than the AMDs; but the AMDs' two additional cores more than make up for the difference. Note that the i5s overclock rather better as I understand it, so if you're planning to try that then it becomes more of a toss-up.)

The 1090T is a very good over-clocker. Mine is running at 3.8GHz without any special cooling and without my spending any significant time tweaking and checking for stability. Reviews suggest that they can reach 4.0GHz with some effort and somewhat higher if you are willing to try really hard. Trying really hard, unfortunately, also takes a significant expenditure.

I second the GTX460 recommendation. Mine is the 768MB version because it was markedly cheaper and more easily available when it was purchased.

Paul

[siegert81]

Some questions:

1. Is the 1 GB GPU necessary or just better than the 768 MB GPU?
2. If I want to focus in on just one smaller k value for proth primes (outside of PrimeGrid, for the purpose of finding Fermat factors), which sieve software should I use and where can I get it?
3. How much faster is ppsieve than NewPGen?

[xilman]

Originally Posted by siegert81

Some questions:

1. Is the 1 GB GPU necessary or just better than the 768 MB GPU?

It is not necessary; it is undoubtedly better at raw computation. Whether it is better by other metrics, such as price./performance, depends on other factors.

Paul

[mdettweiler]

Originally Posted by siegert81

Some questions:

1. Is the 1 GB GPU necessary or just better than the 768 MB GPU?

Significantly better, as I understand it (and representatives of each type are available on Newegg for the same price, so there's no reason to not get the 1 GB version).

2. If I want to focus in on just one smaller k value for proth primes (outside of PrimeGrid, for the purpose of finding Fermat factors), which sieve software should I use and where can I get it?

That would be sr2sieve; it can be found at http://sites.google.com/site/sr2sieve. However, if you are doing just one k (as opposed to a small handful), sr1sieve is even faster. You'll also need the companion srsieve which while slower than either, is needed to initially start the sieve and create a sieve file (since sr1sieve and sr2sieve aren't capable of checking some lower factor candidates).

3. How much faster is ppsieve than NewPGen?

I don't have exact data handy, but it has been determined that sr1sieve/sr2sieve is much faster than NewPGen on individual k's or a handful of them (note that NewPGen can only sieve one k at a time); and for large blocks of k, ppsieve can be as much as a few times faster than sr2sieve. (ppsieve, however, only scales efficiently if you are sieving everything below a particular k-value; for instance, k<10000 as PrimeGrid is doing; there would be no point in doing just k=10000-20000 since it would take just as much time as all k<20000. For this reason, ppsieve is best used in the setting of a large distributed project.)