Given the variety of behaviors and configurations possible with combined CPU & GPU processing, I decided to try a little experiment.

The Machine:
Pentium D830 3.0Ghz dual core (WinXP Pro SP3 32-bit)
ASUS 9600 GSO (600 core, 1700 shader, 1800 memory factory OC)
BOINC 6.3.21 (default settings)

The experiment:
Use a simple program to induce user load on the machine while crunching 2 CPU & 1 GPU unit and see what happens. The program of choice...the standard MineSweeper game included in Windows. First, CPU usage was monitored with nothing actively engaged on the part of the user (just the 2CPU & 1 GPU BOINC load). Next, MineSweeper was engaged in an almost constant usage by repeatedly reloading the game board (i.e., clicking the little smiley-face guy repeatedly).

The Results:
With no load, the CPU usage was typical of that reported by other users with XP in dual core machines. One CPU task (an RCN workunit) at about 48% with the remaining 50% divided such that the GPU work used between 13% and 17% and another CPU task (a PrimeGrid unit) used the rest.

With the induced MineSweeper load, the first task droped from 48% to as low as 22%. The remaining CPU task and GPU task were unaffected.

So, it would appear that the CPU with the GPU assignment is somehow insulated from the user load with other work. Would be interesting to see if this behavior is 1) consistent when testing with other user apps and 2) is the same if a CPU is dedicated to the GPU.

I performed a quick test with the built-in WinRAR benchmark. If it runs multithreaded at normal priority it gets almost 100% of my quad core, if it runs single threaded it gets / takes 25% and the cpu usage of GPU-Grid drops by several percent while the 4 cpu-tasks share the remaining time. After 10-15s the GPU temperature dropped.

So in my case the behaviour is not as clear-cut as yours.

MrS

---

Scanning for our furry friends since Jan 2002

These differences might reflect some of the processor architecture differences, I suppose. The old Pentium D's like mine are basically just two P4's slapped together with separate caches unlike the quads and core 2's. Wonder what the results might be on a new Core2 or on Athllon X2 boxes or a HT P4?

I'd rather attribute it to the scheduler juggling jobs with different priorities around. When I try the minesweeper thing I can generate about a 10% cpu load. It seems like a little of this is chopped off the GPU task, whereas the rest is taken from the regular cpu-jobs. But not very reliable and a bit tiresome ;)

MrS

---

Scanning for our furry friends since Jan 2002