Good cpu/gpu for CudaLucas 2.06

[patgie]

Hi there,

I am really bad at hardware but thinking of putting together GPU for running CudaLucas (so far I have only used prime95).

Could someone please:
a) point to the most basic threads that talk about requirements,
b) is it possible to use 4 (multiple) GPUs through one CPU and if so what are recommendations on CPU in order to handle it?
c) what are performance limiting factors?

Any tips on hardware/manufacturers are much appreciated.

[VBCurtis]

You are more likely to be power-supply limited than CPU-limited if you put multiple GPUs into a single system.

How big is your current power supply? Why do you want to go from no GPU to four in one leap? Buy a GPU. Run it. See what it does to your power bill, room ambient temperature, computer case temperature. If all are quite comfortable, add GPU #2. Repeat until your local power company gives you a call, or you reach the safe limit of your power supply.

Some inexpensive motherboards do not support more than one GPU; you should learn what board you currently have and how many GPUs it can support. You don't need many lanes (x4 is plenty for GPGPU work), but the physical slots have to accept GPU cards.

[kriesel]

CUDALucas means CUDA gpu which means NVIDIA gpu. AMD gpu would mean clLucas or gpuOwL or mfakto.
Used vs. new is up to the user. Unless doing cloud computing, which is a possible way to learn about capabilities at low incremental cost.

CPU capacity is not the limit in running gpu applications. Even CUDAPm1 although it does some operations on a cpu core (stage 1 gcd, stage 2 gcd, gpu load, save checkpoint, etc) the cpu duty cycle is not high. I have a couple gpus running per Core 2 Duo system. I've even run a gpu on an old P4 system on XP. There should be plenty cpu power left for prime95 for nearly any cpu you're likely to have. (Running LL or PRP on igp's though, can seriously impact prime95 performance on the cpu; use mfakto or mfaktc instead there and test.)

Count pcie slots and gpu power cables and consider geometry. Most gpus are two slots wide. If some gpus are low power enough, actively powered pcie extenders can allow use of pcie slots a single slot width apart. Sometimes power cable adapters are needed.

With used workstations, I've found that some configurations that will run initially, later won't. I speculate it's component aging so that power margin disappears. HP Z600s will run usually run 3, but 4 was not stable.

It can be problematic to mix NVIDIA and AMD in the same system. For simplicity I've segregated mine onto separate systems.

Good cooling is critical. So is the ability to pay a significantly increased power bill. See http://www.mersenne.ca/cudalucas.php for comparative first cost, throughput rate, and power per throughput benchmark figures for gpus on CUDALucas, or http://www.mersenne.ca/mfaktc.php for TF. (P-1 relative performance is similar to CUDALucas.)

See http://www.mersenneforum.org/showthread.php?t=23415
and the CUDALucas reference thread etc. at http://www.mersenneforum.org/forumdisplay.php?f=154

Welcome and have fun.

[patgie]

Thank you this is very informative. Yes I would start with one GPU but want to keep an option of scaling up open so want to buy right CPU to start with.

Do I understand correctly that running one GeForce GTX 1080 Ti 24/7 to test 85M exponents would cost roughly $200/£100 a month (0.25$/kWh, 0.13£/kWh)?

[kriesel]

Originally Posted by patgie

Thank you this is very informative. Yes I would start with one GPU but want to keep an option of scaling up open so want to buy right CPU to start with.

Do I understand correctly that running one GeForce GTX 1080 Ti 24/7 to test 85M exponents would cost roughly $200/£100 a month (0.25$/kWh, 0.13£/kWh)?

http://www.mersenne.ca/cudalucas.php find TDP for the gpu.

250w TDP = 0.25kwhr/hr * 720 hr/30-day month at continuous saturation =~ 180 kwhr/month. Then apply the local utility rate; 180 kwhr/mo x 0.25$/kwh = $45/month. (Per gpu)

Units cancel like variables do.
Note, there are reports of some gpus using noticeably more than their nominal TDP when run hard, as in GIMPS computations. Any downstream-of-the-meter distribution loss such as resistive loss in house wiring, power supply inefficiency, etc, and air conditioning or fan power is in addition to the computation above, also. My guess is all that adds up to a factor less than two.

Originally Posted by patgie

Hi there,

I am really bad at hardware but thinking of putting together GPU for running CudaLucas (so far I have only used prime95).

Could someone please:
a) point to the most basic threads that talk about requirements,
b) is it possible to use 4 (multiple) GPUs through one CPU and if so what are recommendations on CPU in order to handle it?
c) what are performance limiting factors?

Any tips on hardware/manufacturers are much appreciated.

I ran a GPU computing system with 10 GPUs and 2 power supplies. For electrical stability I had to split it in 2 separate systems each with 1 power supply and 5 GPUs.


On the big system I used gpuOwl, with both hardware and operating system errors.


On the new split systems I use gpuOwl and Mfakto now. Both systems gained in stability, the errors that affected the big system are gone.


GPU Hardware: Asus Radeon RX580

[Mark Rose]

Cryptocurrency prices have crashed in the last week, especially those mined on GPUs, like Ethereum. Also, new Nvidia cards are being released in a month. You may have the opportunity to pick up a used mining rig for cheap in the next couple of months.