LHC@home

Greetings!

While I'm contributing to other projects,
I thought about buying a computational GPU..



Then I started reading the BOINC, BOINC Trac Wiki, and BOINC Team USA forums.

(More delay of chores while reading...)

Now to decide:
Which project?
Cuda vs Fermi?
16 or 32 bit?
Linux or windows?
Does my PC have enough power and ventilation?
Is my UPS big enough?
Can I commit to the power bill (with cooling the house)?

I would like to hear your comments and advice.


Then, I wanted to try to write some code.
"Just how hard can it be?" My ex-manager's favorite quote.

Oh well, I've done some bit-slice code long ago.....

Any suggestions for that?


Thanks,
Jay

Greetings!

While I'm contributing to other projects,
I thought about buying a computational GPU..



Then I started reading the BOINC, BOINC Trac Wiki, and BOINC Team USA forums.

(More delay of chores while reading...)

Now to decide:
Which project?
Cuda vs Fermi?
16 or 32 bit?
Linux or windows?
Does my PC have enough power and ventilation?
Is my UPS big enough?
Can I commit to the power bill (with cooling the house)?

I would like to hear your comments and advice.


Then, I wanted to try to write some code.
"Just how hard can it be?" My ex-manager's favorite quote.

Oh well, I've done some bit-slice code long ago.....

Any suggestions for that?


Thanks,
Jay

Too many questions with too little info about what you are currently running.

The latest crop of GPUs can be real power hogs so you could easily be looking at needing a new PS. I decided to Upversion to Windows 7x64 and by the time I finished about the only thing that was the same was the case, the keyboard & display.

The GPU I have is Cuda (NVIDIA GeForce GTS 450) which is not anywhere near the top of the GPU crunchers.

SETI does a good job of making sure the Cuda work is made up of short jobs. Keep in mind that running a Task on a GPU is very similar to T4T where BOINC no longer has control over Task swapping on the GPU. Einstein also has Cuda work but their jobs tend to run much longer (at least they did in the past) so they don't/didn't play well with others.

My system is WIndows 7x64, CPU is INTEL i7-2600, NVIDIA GeForce GTS450, 16GB RAM.

Keep the filters clean.

And yes, if you have to up the PS to handle the Cuda & additions to the "box" then you need to evaluate the UPS.

Tom

Tom,
Thanks for the response!!!

I have some older PCs but they have power and CPU speed limits.
(32 and 64 bit, AMD and Intel, Linux and varieties of Windows.)

So, I'm thinking about starting over.
(and cleaning out some old carcasses...)

That's why I appreciate you telling me of your choices.

I guess the deciding factor will be heating up my house and the power bill..

Thanks again,

Jay

(I'll try to change the filter tonight.)

Which project?

If you have NVidia GPU then you can do GPU grid, the other projects SETI, Collatz Conjecture, MilkyWay@home support both.


Cuda vs Fermi?

Cuda is a framework for GPU programing and Fermi is the architecture of a NVidia GPU, if you buy almost any NVidia GPU it will do Cuda.

16 or 32 bit?

No one really uses 16bit, if you have more than 4Gb of ram then you must use 64bit to address all the memory.

Linux or windows?

Your choice, not much difference between the 2.

Does my PC have enough power and ventilation?

Depends on the PSU, what power is it now? Probably have enough ventilation it's cheep to improve that if needed.

Is my UPS big enough?

Depends on how much overage you had when you bought it, I don't worry about UPS for dedicated crunching computers, if the OS is hosed by a crash there isn't much lost.

Can I commit to the power bill (with cooling the house)?

I doubt adding one GPU will impact this much, I had a serious farm once and the cost of the power was the thing that made me give it up, when you have 4-5 high end commuters with multiple GPU that use a lot of power 24/7 running.

I'm currently running Einstein, among other projects. Lately the running time for CUDA w/u's has been about 25 minutes on my machine. And yes, the power consumption and GPU fan noise are non-trivial concerns.

cheers - joe

Win 7 x64
AMD x6 1090T @3.2GHz (water cooled)
NVIDIA GeForce GTX 470
PS 850W (I think)



[/img]

The problem with the heat from a GPU is like any other problem: if you solve it the stupid way then you'll not be happy. If you think about the problem instead of just rushing out and buying the latest trendy gizmo solution then you'll be much happier. The stupid way to solve the heat problem is to let the hot air from the GPU mix into the room air and heat the room up and then try to cool the GPU with that hot air. Think about it, does trying to cool a hot thing by blowing hot air at it make any sense? If you think it does then you're lost already.

Of course you can turn on air conditioning to cool the hot air but that costs money. If you didn't have to spend that money on AC you could buy a better GPU.

So think.... how can you prevent the hot air from the GPU from heating up the room? In the winter maybe it's not a problem. In the summer, however, position the computer below a window, build a duct that prevents the hot air from mixing into the room and causes it to travel up to the window where you have a fan that sucks the hot air in and blows it out the window. You can make a duct out of cardboard and tape or you can buy cheap 4" (100mm) aluminum clothes dryer duct which you can cut with a good pair of scissors. Or you can raise the computer up on a high table and position it directly in front of the fan intake which eliminates the need for a duct. Very simple, very inexpensive, works very, very well.

Another solution is liquid cooling but run the coolant lines through the wall and place the radiator and fan outside. Then the heat doesn't stay in the house.

Anyone who claims there isn't much difference between Linux and Windows is a few bricks shy of a load. One big difference is that Linux works while Windows is garbage. The other major difference is that Linux is free whereas Windows is not and you keep paying and paying and paying and paying.... for more garbage. Use Linux and take the money you save and buy a better GPU. AMD and nVIDIA both have good drivers for Linux, Ubuntu is almost to the point where you can install them from repositories and in fact they may actually be there now.

What you won't find with Linux is GUI apps like GPU-Z but there are CLI apps for monitoring and setting up the GPU, GUI apps are on the way.

Contrary to what Tom says, the real power hogs are the older GPUs, if you consider bang for your buck in other words flops per dollar. The new GPUs use 22nm technology which means the transistors are smaller which in turn means they use less power per operation. On the other hand, because the transistors are smaller they cram more of them into the same size GPU and the thing ends up using more power but it does far more work with the same amount of power. So get a new GPU... an nVIDIA 600 series or AMD 7900 series and avoid the older ones, talk to the guys at GPUgrid and get their advice, they'll tell you that you cannot afford to run an old GPU. Right now the nVIDIA GTX-660Ti is the sweet spot... excellent performance at reasonable price.

Another reason for new GPU... GPUs are evolving very rapidly and older models are becoming obsolete. Already at GPUgrid a number of older nVIDIA cards are no longer usable. nVIDIA 4xx cards are on their way out at GPUgrid and are capable of running only the short version of the tasks, not the long version.

ATI vs. nVIDIA....

nVIDIA uses CUDA and OpenCL, AMD uses OpenCL. OpenCL is a general, non-optimized platform, CUDA is highly optimized for computation intensive applications like you find in BOINC world. The consensus in the general community and not just the BOINC community is that CUDA outperforms OpenCL for sheer computing power. AMD owners don't like to admit that but they are biased so look at unbiased reports from outside the community. Some projects have OpenCL apps for both AMD and nVIDIA, others have CUDA for nVIDIA and OpenCL for AMD. If you buy an nVIDIA based card you'll want to run the projects that use CUDA because in general and over the long run those will drive your card the hardest and use it most efficiently.

I have had an nVIDIA GTX-570 for over a year and recently bought an AMD 7970 which is the latest and best AMD has to offer. It was a mistake. I should have bought an nVIDIA instead. The number of projects offering OpenCL apps is dwindling. Poem has a good AMD app but their tasks are down to a trickle. Milkyway has an AMD app but I would not crunch that project if it was the last project standing, it's a rogue project run by misfits. The prime number projects have AMD apps but the biggest use for prime numbers and rainbow tables is for the banks to secure their transactions and given the fact they they have turned out to be the biggest crooks on the planet (read the latest scandal in the new: mortgage rip-offs, drug money laundering, trucking with terrorists) they can go pound sand, they take enough from my pockets already. WCG has an AMD app for their "help cure cancer" project but that will be finished soon. For me the sensible way is nVIDIA because if a project has a CUDA app then it's right up nVIDIA's alley and if they only have an OpenCL then nVIDIA will work just as good as AMD with that.

I don't plan on buying anymore high end CPUs. My strategy from here on is to buy high end GPUs plus just enough CPU and motherboard to drive the GPU. Yes, that is a consideration... your CPU and motherboard must be fast enough to feed the GPU the data it is to compute and remove the results of the computations and store it to disk or whatever. If your CPU and motherboard aren't fast enough then the GPU will not run optimally. The latest batch of GPUs use PCI-E x16 3.0 which moves data across the PCI bus much faster than PCI-E x16 2.0. If you're going to get a new motherboard and a new generation GPU then you would be wise to get PCI-E x16 3.0. Not every app can use it at this point in time but you can be sure that in the future you will want to have it.

So my next system will be a mobo with 4 PCI-E x16 3.0 slots that run at x16 speed even if all slots are occupied (some mobos drop x16 back to x8 if more than 1 slot is occupied) and a fast 4 core CPU to feed the 4 GPUs I will put onto the mobo. I'm not sure where I'll find a power supply big enough for 4 high end GPUs but I will build one or modify off the shelf PSUs if I need to. The UPS?... any UPS you buy off of a shelf is junk (because the battery sits for months/years in a warehouse with no voltage applied to it so it discharges which causes the sulphur in the acid to deposit onto the plates in the battery and then it's toast) so I have my own custom UPS which works very well. Don't buy a UPS from a store, waste of money.

Many Thanks.
I had read the responses - but just now responding.
T H A N K S!!!!

Jay