This situation happens *only* when there are two different GPUs with *significantly* different crunching power. BOINC uses the fastest GPU for calculating work requests x the number of GPUs. So for example, assume you have a 9500 GT and a GTX 580. BOINC would automatically calculate work fetch based on the 580, times two cards. That could end up downloading so much work that it cannot return it all in time, depending on the size of your cache and the deadlines of the project. Because the 9500 is not really a 580. So by default, if BOINC detects a large enough difference between the two (or however many) GPUs, it will ignore one of them. I forget which, the faster or slower. Anyway, if you want, you can over ride this behavior with the method you describe. But understand that you are taking a risk of over-downloading tasks, and being late. So be sure to keep an eye on your queue, and keep it small until you have a good handle on the right cache size to keep without issues.