Here is my conclusion after I've run GPUGrid for a few hours. I have previously run another non-Boinc project, Folding@Home GPU, for a couple of days as well.

For all GPU-based crunchers the problem is lack of priority queues. A GPU cruncher will hog all available GPU time, leaving for example desktop unresponsive (scrolling a FireFox window, dragging a window around would be jerky). This was not a problem before, when only one application (desktop or game) accessed the GPU resources at any given time. It is a problem with CUDA. nVidia driver MUST implement a way to prioritise tasks on GPU. So that, for example, GPUGrid would run at lowest priority and yield GPU to a high-priority application (a game) or to a real-time application (desktop). Really the same way it is done on a conventional CPU by a modern OS.

BOINC-specifically, GPUGrid is actually wasting CPU time. I have a 4-core CPU. Normally all 4 cores are kept busy by various BOINC tasks. In comes GPUGrid. It is counted as a task, but as it runs on GPU, one of the cores of my CPU is not utilised. I tried setting the limit in preferences to allow 5 simultaneous tasks on a multi-core CPU, but BOINC would still only start 4 and with GPUGrid almost not touching the CPU, it is only used for 75%.

One other note, memory is important. I have 1G or better on all GPU cards and see little of the stuttering and the like on any system. Of course I do not do high intensity gaming ... but even on the 9800 GT I have no hesitations on the system I can point to the GPU as the cause ...

YMMV

You can have 4+1 on a quad core machine (assuming you have 1 gpu). Under 6.4.5 you can fudge it my overriding the number of cpu's, ...

Just use 6.5.0 and running 4+1 shouldn't be a problem.

And what you're generally asking for will be delivered with DirectX 11, as far as I understand. It introduces compute shaders, which can run along whatever else is going on on the GPU.

Paul, it's not the memory (yet). With the first WUs GPU-Grid plus an XP desktop consumed about 70 MB of video memory. Some of the current WUs are larger, but not by much (see proportional increase in time per step for these). Your not seeing this drastic slow down in screen response time because your cards are too fast. Your 9800GT is a good 4 times faster than his card.

MrS

as soon as someone mentions a card slower than a 9600GT, it's just not worth to keep these in mind!

I also think, that this "all or nothing" GPU usage is just too restrictive. I know, my graphics card might not be "fast enough" for GPUGRID, but it is able to finish GPUGRID WUs in time (GeForce 9400GT, 512 MB). The only problem is, that I can run 11 other BOINC projects (CPU) with 100% CPU usage without any problems in the background. But GPUGRID is just sucking too much from the GPU resulting in sluggish and joppy desktop usage which is no fun.

I often read that the SETI@home GPU version doesn't have this problem (can't check as I have Linux and there is no CUDA version for Linux out yet). So, it doesn't seem to be a CUDA API problem.
____________
GPUGRID on 64 Bit Linux with GeForce 9400GT (512MB), AMD X2 4850e (2 * 2,5 GHz), 8 GB RAM

Okay, this means it's a problem of the GPUGRID WU's. I don't know the algorithms of WU's but I'm a software developer and therefore can't see a problem to just design the WU processing in a more resource friendly way. I guess, all BOINC projects have some calculations within loops and countless iterations. So, it would be no problem to split such calculations in smaller "chunks" ... just like SETI@home and Folding@home?
____________
GPUGRID on 64 Bit Linux with GeForce 9400GT (512MB), AMD X2 4850e (2 * 2,5 GHz), 8 GB RAM