Folding Forum

GreyWhiskers wrote:

Now after this job is done I hope I can get into this and go buckwild maybe get a job a day done would be nice concidering Im not real sure how the points thing works with evga and how they do their rewards program but Im sure it takes a lot of points lol .

I wouldn't be surprised if you don't finish a typical Fermi GPU Work Unit in, say, 3.5 hours or even faster. If you only have the GPU client configured at night, you should get multiple of these WUs done. Could be faster, depending on your stable overclock.

OKay someone wrote above that I should put the gpu in a slot at night.... does that mean the cpu and gpu will be working together on a job? or does that mean the gpu will work alone? is one faster then the other?

They are completely independent, each getting its own series of work units. The SMP client uses CPU cycles on, typically, large protein molecules - the largest these days for 12 CPU core monsters have several million atom proteins. The GPU work units are typically simulation of folding of proteins with several hundred atoms - but they process exceedingly fast on the Fermi GPUs like your GTX 460. The vast majority of the processing is done in the GPU itself, but a small amount of CPU time is used to move data in and out of the GPU.

They can typically coexist well, especially if you have left a core or so off of the SMP client for other things (I think that using the -SMP 3 flag was suggested at one point in this thread).

I myself run an Intel Sandy Bridge i7-2600K under Windows with one SMP client and one GPU client (a GTX 560 Ti). I experiment from time to time how to configure the SMP (e.g., the -SMP 8 or -SMP 7 or -SMP 6) to coexist with whatever other loads are going on in the CPU.

There isn't a cookie cutter answer - so some experimentation, looking at how the "TPF" or Time per Frame translates into points per day to generate the fastest return for your work units.

EDIT: re: coexistence. I just leave both the SMP and GPU clients running 24/7. If you must, you can shut down the GPU folding client during game play - but again, you can to experiment to see if you need to.

or does that mean the gpu will work alone? is one faster then the other?

Let's say they are different.

Most of the SMP work units I get are the "bigadv" types that use all 8 threads of the i7 2600K Hyperthreaded CPU. The work units are typically on one-million atom proteins, and complete in 48 hours plus, although one project of work units, P2684, takes about 1/3 more time.

Most of the GPU work units we get these days take my GTX 560 Ti (overclocked to 950 MHz) an exceedingly consistent 2 hours 17 minutes. There are a series of smaller work units, though, that process in one hour and 11 minutes.

So, in V7, the client keeps track of the cores - fetches new work when needed and sends completed work back to the Stanford servers. There is an asynchronous fetching and sending for however many cores you are running.

thanks a lot this is great information I will do experimenting and see how fast I can get this bugger going. might even over clock for the first time since the mobo has a easy overclocking switch on it I have been itching to try .

With nvidia I wouldn't recommend -smp 3, his gtx460 might need some more feeding then my gtx280 but it would still be very very little in total.

If you start oc'ing the cpu/ram, run stability tests such as stresscpu v2 and orthos 2004, stresscpu will test your cpu and orthos will test your memory.

well the first night wasnt bad I guess a total of 5 jobs completed all on my gpu the cpu job only made it halfway through the job I was a little surprised by that

As I said, each project, each work unit, will have a certain track record of how long it takes in a given environment. For me, the fastest GPU work units take just over an hour; the slowest SMP work units take 3 days; and some recent CPU work units I run on an old uniprocessor incapable of SMP were running over 7 days; the GPU work units on my older AGP bus ATI HD4670 take about 11 hours. And, everything in-between.

I think there are some databases around for expected times to complete the different work units, but there is so much variability to hardware and environment, that you may want to do it yourself as you go along. There are third party tools, e.g., HFM.net, that will keep stats like that for you in the old v6. There has been considerable discussion that when Stanford's new v7 client comes out of beta and has its data structures and APIs stabilized that third party sites will provide such new tools for v7.

If you haven't looked, there is a link at the very top of the forum page called "tools" that takes you to a list of various tools that can help you set up, monitor, evaluate your hardware, etc. None of them yet are specific to v7. While they don't keep track of individual work units, third party stats sites continually analyze a monster flat file that Stanford provides each hour and makes the analysis available to individual users and teams. I frequently look at Kakao and XCPUs web sites - links are in the tools link.

The big caveat seems always to be, "YMMV".

EDIT: here's a snippet of a couple of days of production for my Sandy Bridge system - you can see the wide variety of work units, the "Frame Time", which is the time to complete each 1% frame, and the points/points per day. I still run this one computer in v6 so I can continue to get these stats from the HFM.net tool. I run v7 on my smaller systems. [sarcasm on] I remind myself of the movie "Men who stare at goats" as I ponder these lists... [/sarcasm off]