mfaktc: a CUDA program for Mersenne prefactoring

[James Heinrich]

Originally Posted by mdettweiler

might adjusting SievePrimes help reduce the bottleneck?

Just let SievePrimes self-adjust to achieve minimal bottleneck.

If you can't get >90% GPU usage with a single instance, start up a second instance of mfaktc and you should get close to full GPU usage, with CPU load spread across multiple cores (and with the additional CPU power, SievePrimes should self-adjust higher than with a single instance, again giving overall better throughput).

[Ungelovende]

Originally Posted by Xyzzy

Is it difficult (heat, physical space, power, 16x slots) to get them in the case?

This is my setup:
Antec Performance One Series P193
MSI P67A-GD55 (B3) LGA 1155 Intel P67
Cheap cpu-cooler
ZOTAC ZT-50201-10P GeForce GTX 570
ASUS ENGTX570/2DI/1280MD5 GeForce GTX 570
Chieftec Nitro Series BPS-1200 1200W PSU
+ cheap RAM
+ Sandy 2600k

The temperatures:
http://img339.imageshack.us/i/tempsandy.png/

Difficult? I needed a beer and a screwdriver

Is running them SLI an option?

Not shure. This computer is not used for gaming - crunching only. Never testet SLI.

The CPU or the GPU?

The CPU (i7-920) is too slow to feed two GTX570

[mdettweiler]

Originally Posted by James Heinrich

Just let SievePrimes self-adjust to achieve minimal bottleneck.

If you can't get >90% GPU usage with a single instance, start up a second instance of mfaktc and you should get close to full GPU usage, with CPU load spread across multiple cores (and with the additional CPU power, SievePrimes should self-adjust higher than with a single instance, again giving overall better throughput).

Aha, so that's how it works. That would explain why it takes a little while upon initial startup each time for M/s to peak.

Looking at the mfaktc console output, I see it is actually using SievePrimes=5000 at optimal speed. I wasn't paying attention to that figure on the console earlier, but presumably it was trying different values when the speed was initially lower.

Thanks!

[TheJudger]

Hi all!

Originally Posted by nucleon

Agreed - in time mfaktc needs to get automated submit/get work. Or better yet - integrated into prime95. Prime95 handles the submit/get work, and has a generic option to run gpu code as separate process.

But Oliver is but one person. I think we can handle the inconvenience of manual submit/get work for the moment while Oliver works on his priorities. Getting version 0.17 out and getting the code optimal/valid.

My priorities? Currently mfaktc has a low priority...
0.17 coding should be finished but I want to edit the README a little bit.

Originally Posted by Xyzzy

Random, late night questions:
2 - Given most of us run 2 instances, would an i3 be enough? One of them has the clockspeed and HT. Does HT make a difference? Does reduced L3 cache make a difference?

i3 should be OK for a single 570.

Originally Posted by Xyzzy

3 - Can you run 2 GPU cards and use them both? Is it SLI or separate?

Yepp, multiple GPUs are supported for a long time. But you can't use multiple GPUs for one instance. Each GPU needs at least on instance of mfaktc.

Originally Posted by Xyzzy

4 - How much more productive would a 580 be over a 570, noting it is ~$150 more?

Comparing 2 GPUs of the same compute capability is very easy. Relative raw GPU performance: (512cores * 1544MHz) / (480cores * 1464MHz) = ~1.125
So the RAW performance of a GTX 580 is equivalent to 1.125 GTX 570. Usually this means a little bit less sieving on the CPU so I would expect something like 10% more throughput.

Originally Posted by Xyzzy

Our 4 non-GPU quads will be retired within the next month.

Are they worth running LL and/or P-1 on CPU?

Originally Posted by mdettweiler

Hi guys,

Recently, I've been playing around a bit with mfaktc on a GTX 460 (768MB) running on a system with a stock Q6600 CPU. I'm currently taking M332228447 from 75 to 81 bits (79>80 in progress right now). With the default SievePrimes=25000, I'm getting speeds of about 109M/s. The CPU appears to be the bottleneck, with mfaktc using 100% of one core and 82% of the GPU. Does this sound like what I should be optimally getting on this GPU/CPU/exponent combination, or might adjusting SievePrimes help reduce the bottleneck?

Also: I've noticed that whenever I have to stop and restart mfaktc, when it first comes back it starts at only ~75M/s, and takes about 1-2 hours to work its way back up to the usual ~109M/s. Is this normal behavior? Does anyone know why it does this?

Yepp, a single core of your CPU is the bottleneck.
The behavior you've noticed the the adjustment of SievePrimes. Each time you restart mfaktc SievePrimes starts at 25000 (default configuration). On your setup I assume the it goes down to 5000 after a while. You can either add another core (start a second instance of mfaktc working on an other exponent) or set SievePrimes to 5000 in the mfaktc.ini to avoid the behavior (but still run CPU limited).

Btw. the raw GPU speed is not the perfect measurement for performance, take a look the the per-class-runtime.


Oliver

[Karl M Johnson]

Originally Posted by Xyzzy

Is running them SLI an option?

From what I've heard, having two GPUs connected in SLI will not hurt compute applications.

[Xyzzy]

Yepp, multiple GPUs are supported for a long time. But you can't use multiple GPUs for one instance. Each GPU needs at least on instance of mfaktc.

Do you have to assign each instance to a particular GPU or do the cards act like one big GPU bucket?

Are they worth running LL and/or P-1 on CPU?

They are doing P-1 right now. Two of the boxes have 16GiB of RAM, one has 8GiB and one has 4GiB. They use almost the same amount of electricity as our two GPU boxes, so if we get rid of them we can maybe justify two more. They are OEM HP boxes so putting a GPU in them is pushing it thermally, physically and electrically.

[Xyzzy]

This is my setup:

Thanks! The Newegg links are useful too!

We are trying to figure out if there is a motherboard (maybe an ASUS) that will take 2 of the 3 slot Asus 570 cards like we have now. We are happy with them so we want to stick with them.

[steinrar]

Xyzzy; I own that type of GPU you link to, but don't like it. It heat up other component inside unless you have enough fans pushing air into computer giving posetive pressure inside. I have a similar setup like Ungelovende with 2x570 cards that push most of the air out of the pc, much better I think. My 5 cent...

[Christenson]

Originally Posted by TheJudger

Hi all!

My priorities? Currently mfaktc has a low priority...
0.17 coding should be finished but I want to edit the README a little bit.

Oliver

I have just asked if I should be Oliver's assistant and add the automatic assignments part to the code, copying of as much as possible from Prime95/mprime.

I hope to be the second body, and learn enough to contribute to the core code and/or algorithms at some point.

Eric Christenson

[Xyzzy]

Xyzzy; I own that type of GPU you link to, but don't like it. It heat up other component inside unless you have enough fans pushing air into computer giving posetive pressure inside.

We have a big fan in the front of the case and a big fan on the ceiling of the case, so the air crosses the card, moves up to the CPU and out the top. The PSU heat is pushed out the bottom of the back. Our GPU temps under load rarely exceed 65°C, but we are not overclocking. (Even the "stock" overclock is mild.)

We would prefer a slimmer GPU but running one 24×7 under full load has to cause higher temps and lower life, right? We doubt these GPU cards were designed for a 100% duty cycle.

[Christenson]

30 Hour validation test result:
no factor for M53953421 from 2^50 to 2^66 [mfaktc 0.16p1 75bit_mul32]
no factor for M53953421 from 2^66 to 2^67 [mfaktc 0.16p1 barrett79_mul32]
no factor for M53953421 from 2^67 to 2^68 [mfaktc 0.16p1 barrett79_mul32]
no factor for M53953421 from 2^68 to 2^69 [mfaktc 0.16p1 barrett79_mul32]
no factor for M53953421 from 2^69 to 2^70 [mfaktc 0.16p1 barrett79_mul32]
no factor for M53953421 from 2^70 to 2^71 [mfaktc 0.16p1 barrett79_mul32]
no factor for M53953421 from 2^71 to 2^72 [mfaktc 0.16p1 barrett79_mul32]
no factor for M53953421 from 2^72 to 2^73 [mfaktc 0.16p1 barrett79_mul32]
M53953421 has a factor: 16867347823849190640239
found 1 factor(s) for M53953421 from 2^73 to 2^74 [mfaktc 0.16p1 barrett79_mul32]
(This factor was known from my P-1 effort)(Is it worth telling Primenet I duplicated the work?)
no factor for M3321934241 from 2^76 to 2^77 [mfaktc 0.16p1 barrett79_mul32] (took 3 hours or so)
(Operation Billion Digits...I'll take that up to 2^82 or so in the next two weeks)

Noting that there are stock fan kits for GPUs -- in my personal experience, the fans die first if the chips are kept cool. Hard drives die. But at only $100 for the GTX440, I can't complain.... and with the temperature monitoring and throttling, they are designed to support hours at a time of full-throttle gaming.