System Recommendation
 

[SIZE=3][FONT=Calibri]I’m looking to have a system built for me, dedicated to this effort, and I’d like some input for the hardware specs.[/FONT][/SIZE]

[SIZE=3][FONT=Calibri]I want to do some GPU computing, along with 1 or 2 cores doing P-1’s, and the rest doing either LL’s or LL-D’s.[/FONT][/SIZE]

[SIZE=3][FONT=Calibri]I’m looking at a MB with 2 CPU sockets on it, and for my budget, I have a choice of installing (1) - Intel Xeon DP Six Core L5640 2.26Ghz 5.86GT/s 12mb 60w 1066Mhz, or (2) - Intel Xeon DP Quad Core L5520 2.26Ghz 5.86GT/s 8mb 60w 1066Mhz. Any suggestions as to which configuration will give me better performance?[/FONT][/SIZE]

[SIZE=3][FONT=Calibri]I’m also looking to install about 8GB of Memory – DIMM 4096mb 1333 MHz Registered ECC DDR3 on each CPU if I go with the Quads, or DIMM 8192mb 1333 MHz Registered ECC DDR3 if I go with the Six Core … any opinions here?[/FONT][/SIZE]

[SIZE=3][FONT=Calibri]Finally, which NV graphics card would you recommend for this config?[/FONT][/SIZE]

[SIZE=3][FONT=Calibri]Thanks,[/FONT][/SIZE]
[SIZE=3][FONT=Calibri]Doug[/FONT][/SIZE]

Is 2 pcs not an option? That would be much cheaper as you could use desktop parts(and faster because you would get better parts).

2 PC's is an option, just thought I'd get some efficiencies combining them. Have a recommendation on a CPU, Memory, and GPU?

Thanks,

GTX 480/GTX 580 with best factory OC is your choice.

For a long time it's been better to get two PCs than a single two-socket PC; basically, the period in which the two-socket machine was unacceptably slowed down by the shared memory bus ended and was immediately followed by the period in which Intel sold CPUs for two-socket PCs at an inordinate premium.

(I do have a two-socket machine, a dual quad-core Opteron, because someone this time last year was selling 4GB DDR2 modules at £40 each on ebay and two quad-core Opterons at £95 each; with that sort of discount it seemed just about worth paying the full £350 price for a motherboard that could use them and £70 for the slightly exotic power supply it needed).

So: first, wait until Intel have sorted out the glitch with Sandy Bridge chipsets. Then for CPUs use Sandy Bridge i7 2600 (2600K if you like overclocking) to get the new architecture, high speed and hyperthreading; stick two 4G modules with each CPU; I would be tempted to use the on-CPU video on one of the PCs, on an H67 board, and stick a GeForce Ti 560 in the other on a P67 board. You'll need about a 600W power supply for the P67+560, 300W will be more than enough for the H67.

[QUOTE=fivemack;251163]Then for CPUs use Sandy Bridge i7 2600 (2600K if you like overclocking) to get the new architecture, high speed and hyperthreading ... and stick a GeForce Ti 560 in the other on a P67 board.[/QUOTE]

If you're goal is prime95 computing use a i7 2500K with no hyperthreading. Some apps can make use of hyperthreading, prime95 is not one of them.

If you're cost conscious, consider a GTX 460 board. The 560 board is faster, but the 460 board is significantly cheaper.

A agree with all previous posters, Xeon is not the way to go.

My goal is definitely prime95 computing, been doing it for almost 2 yrs now, and I'm addicted.

My next questions is which OS? 3 of my machines are running Win7-64. I have a little experience with *nix, but it's been many years. Is there that much performance difference between them?

Final question is, do all the programs showing up now, to utilize GPU's, require manual interaction with PrimeNet to get assignments, and if so, any projection as to when it will be supported in a similar manner as CPU's?

Thanks for all the suggestions. I'll start my research and shopping now. Keep up the good work ...

Doug

[QUOTE=drh;251195]My next questions is which OS?

Final question is, do all the programs showing up now, to utilize GPU's, require manual interaction with PrimeNet to get assignments, and if so, any projection as to when it will be supported in a similar manner as CPU's?[/QUOTE]

OS is irrelevant to prime95, but do select the 64-bit version. Linux will save money, but use whatever makes you most comfortable.

GPU programs are all manual right now. It will be a long time before these are incorporated into prime95. It isn't hard to reserve a lot of exponents manually, so it should be possible to set up a process where you only have to tend to the client once every few months.

[QUOTE=Prime95;251193]If you're goal is prime95 computing use a i7 2500K with no hyperthreading.[/QUOTE]

Quick clarification - Should it be an i5 2500k or an i7 2600k?

[QUOTE=drh;251479]Quick clarification - Should it be an i5 2500k or an i7 2600k?[/QUOTE]
I would guess the i5-2500K since no i7's come with hyperthreading.

[QUOTE=drh;251479]Quick clarification - Should it be an i5 2500k or an i7 2600k?[/QUOTE]

My bad, I meant i5 2500K. This chip has no hyperthreading.

The i7 2600K does have hyperthreading, but costs a bit more.

Thanks ... back to shopping!

Thought I'd give an update, or rather the results of my system. It's been up and running for almost 2 weeks now. It is an i5-2500K, 560Ti Nvidia card, 8Gig of RAM, and no OC at this time.

I have 2 instances of mfaktc running, using 2 cores, 1 core doing P-1's, and 1 core doing LL's. Current P-1's average about 19 hrs to complete, LL's only take about 0.034 sec per iteration, and mfaktc is producing about 190 GHz-days/day.

I had the system built for me. It's super quiet with no special noise reduction, can't even hear it running, and doesn't seem to be putting out very much heat. I'm very impressed so far.

So the 560Ti Nvidia card is quiet...the GT440 is noisy, can hear the fan in the next room, gets about 90GHz days/day. I'm chasing automating mfaktc a bit more....

[QUOTE=Christenson;261747]So the 560Ti Nvidia card is quiet...the GT440 is noisy, can hear the fan in the next room, gets about 90GHz days/day. I'm chasing automating mfaktc a bit more....[/QUOTE]

It's super quiet, almost silent. I have to keep looking at the LED's or feel the intake or exhaust vents to verify that it is really running. Here is a screenshot of GPU-z with 2 instances of mafktc running.

Doug

If you click the 95% figure in gpuz it changes from min/max/average. To be accurate you need to start gpu-z after mfaktc has been running for a bit (say 1min or so).

If the average is 95%, then you might be able to wring a couple more percentage points by hard coding the sieve primes value in the .ini file.

If you hard code both instances to say 23000. And then check cpu-z or your completion times to see if it improved things.

-- Craig

[QUOTE=nucleon;261804]If you click the 95% figure in gpuz it changes from min/max/average. To be accurate you need to start gpu-z after mfaktc has been running for a bit (say 1min or so).

If the average is 95%, then you might be able to wring a couple more percentage points by hard coding the sieve primes value in the .ini file.

If you hard code both instances to say 23000. And then check cpu-z or your completion times to see if it improved things.

-- Craig[/QUOTE]

Thanks for the tip.

Actually the min/max/avg is 87/95/92. With an average of 92 do you still suggest that I use 23000? Don't I also need to change SievePrimesAdjust=1 to something else, comment it out, or leave it alone?

Doug

[QUOTE=drh;261805]Thanks for the tip.

Actually the min/max/avg is 87/95/92. With an average of 92 do you still suggest that I use 23000? Don't I also need to change SievePrimesAdjust=1 to something else, comment it out, or leave it alone?

Doug[/QUOTE]
Change it to 0 I think.

[QUOTE=drh;261805]Thanks for the tip.

Actually the min/max/avg is 87/95/92. With an average of 92 do you still suggest that I use 23000? Don't I also need to change SievePrimesAdjust=1 to something else, comment it out, or leave it alone?

Doug[/QUOTE]

Cool, you should get about 5% more with hard coding it. You need "SievePrimesAdjust=0" if you want to manually adjust the figure.

Adjusting these settings is more art than science.

You want to set it such that your CPU is underused rather than the GPU being underused.

My plan:

1)
Start back down to 5k. This is the minimum the CPU can do. Make sure you do hit 99% average or thereabouts.

2a)
If you don't hit 97-99% run up another instance if you have the core for it. If another core (@5k sieve primes) doesn't increase GPU usage, then there is a system limitation that you can't get around. (Not enough I/O bandwidth etc...)

2b)
If you're hitting 99% at 5k sieve primes. This means it's possible to hit 99% GPU usage at higher values. What you want to achieve is what's the max you can set it to and still be at 99%.

Given your figures so far, I'd say 20,000 is probably a good value to start with. But again this process is only useful if you're consistently doing the same work.

The optimum sieve primes value can change depending on the bit depth and the exponent size. I find lower bit depths are harder to optimize.

It's up to you if you feel time to muck around with it is worth another 5% of results.


-- Craig

[QUOTE=nucleon;261822]Cool, you should get about 5% more with hard coding it. You need "SievePrimesAdjust=0" if you want to manually adjust the figure.

Adjusting these settings is more art than science.

You want to set it such that your CPU is underused rather than the GPU being underused.

My plan:

1)
Start back down to 5k. This is the minimum the CPU can do. Make sure you do hit 99% average or thereabouts.

2a)
If you don't hit 97-99% run up another instance if you have the core for it. If another core (@5k sieve primes) doesn't increase GPU usage, then there is a system limitation that you can't get around. (Not enough I/O bandwidth etc...)

2b)
If you're hitting 99% at 5k sieve primes. This means it's possible to hit 99% GPU usage at higher values. What you want to achieve is what's the max you can set it to and still be at 99%.

Given your figures so far, I'd say 20,000 is probably a good value to start with. But again this process is only useful if you're consistently doing the same work.

The optimum sieve primes value can change depending on the bit depth and the exponent size. I find lower bit depths are harder to optimize.

It's up to you if you feel time to muck around with it is worth another 5% of results.


-- Craig[/QUOTE]

So, here's what I've found out -

When I set SievePrimes=5000, although the "avg. rate" increases to about 119M/s, the "candidates", "time", "ETA", and "avg. wait" also increase, which increases the total amount of time it takes to complete a test. The GPU will max out at 98%, and the CPU is at about 99%.

I get the same type of results setting SievePrimes=20000, just with a little less increase in each column.

BTW, I'm currently running M76,xxx,xxx from 69 to 73 bits.

I'm thinking I may be better leaving it set up "default" or is it worth the extra time for the increase in "avg. rate"?

[QUOTE=drh;261861]I'm thinking I may be better leaving it set up "default" or is it worth the extra time for the increase in "avg. rate"?[/QUOTE]

Total runtime (and time per class) matters as long as you don't aim for maximum heat generated on GPU :wink:

Oliver

[QUOTE=drh;261861]So, here's what I've found out -

When I set SievePrimes=5000, although the "avg. rate" increases to about 119M/s, the "candidates", "time", "ETA", and "avg. wait" also increase, which increases the total amount of time it takes to complete a test. The GPU will max out at 98%, and the CPU is at about 99%.
[/quote]

Okay, this means the maximum we can drive your GPU is 98% in your system.

[QUOTE=drh;261861]
I get the same type of results setting SievePrimes=20000, just with a little less increase in each column.
[/QUOTE]

Awesome, if GPU is averaging around 98% still, then sieveprimes=20k looks like optimum solution for you.

Looking at what you've said (and _if_ GPU is at 98% with sieveprimes=20k), then hardcoding has yielded around 6.5% more usage on the GPU.

[QUOTE=drh;261861]
BTW, I'm currently running M76,xxx,xxx from 69 to 73 bits.[/QUOTE]

This is good. Hardcoding sieveprimes is only useful if you have predictable loads.

[QUOTE=drh;261861]
I'm thinking I may be better leaving it set up "default" or is it worth the extra time for the increase in "avg. rate"?[/QUOTE]

We're talking about optimizing here. If you're not comfortable and can't sleep at night by all means leave it at the defaults. Run it hard coded for a day or two, check completion times and see if there's a difference.

-- Craig