GPU Price Performance Curve

I used to run 3 x GTX 460 OC... and a GTX 480. But UK energy prices have hit me hard, so I stopped.

I imagine the throughput of my collection of PC's is still on this site.

dunx

But with that one pc you still do more than the average member here with several pcs dunx!

(I need a rainy day so I can install the nVidia card I just got)

I just stopped by here to look at those charts for some friends.

.....back to work.....here at home of course.....

I added a GTX 570 to my Core i7 together with the existing 9500GT. To my dismay, my GTX 570 usage as shown by the GPU Meter (Gadget) on win 7 (64bit) never crosses 60% or so. My 9500GT operates above 80%. Is all this what the GPU price performance curves (of Robert) is talking about?

Can any one help me to get the maximum credits out of my investment. Normal tweakings on BOINC manager (6.12.34(x64)) are done.

Thanking in advance.

Abhijit

Two possible improvements you might consider are to use the anonymous platform mechanism (app_info.xml) to run two or three BRP4 tasks simulataneously, and to adjust the CPU configuration with a view to improving the servicing of the GPU.

There is a considerable amount of prior posting here on the app_info multiple task, matter, so I'll just mention that if you lack experience at this, you may well make a mistake and lose your queue of work at first try, but if you start from a recently posted model, and check each entry against what you can find in your own client_state.xml (making sure, for example, that file names and application names match exactly) you should be able to get there. You'll also probably find you want to use the specification for both your GPU and your CPU projects, which if done right can help avoid some of the more severe scheduler oddities but is rather a pain in the neck.

Regarding CPU--the issue is that the GPU task has to sit around and wait when it wants services from the support application (*cuda32.exe) running on the CPU. So, especially if you get up to 2-fold or 3-fold, you want somehow to assure that that (those) programs run when wanted as promptly as you can manage. Many people advocate using the preference item "On multiprocessors, use at most % of processors" specification to hold down CPU usage by their CPU BOINC applications, thus "saving room" for the GPU support. I've had some luck doing both some of that, and raising the priority of the cuda32 applications (using Process Lasso), and, on my hyperthreaded CPU, segregating the constantly running GW CPU applications and the intermittently running--but urgently needing prompt service cuda32 applications to opposite virtual halves of the physical cores.

Of course, anything which makes your CPU run faster, especially memory operations, will help also, so you could consider trying more aggressive memory timings, CPU overclocking, and the like if that suits your style.

As I've reported at length in my GTX 460 thread, I've recently been through this sort of tuning exercise for two GTX 460 hosts, and got to roughly the mid-80% GPU load as reported by GPU-Z running 3-fold with raised priority on the cuda32 processes on two hosts, one of which was only about 58% GPU load running 1-fold without tinkering.