I propose a plan that when the WUs ready to send is 0 and there are less than 3000 active WUs you implement the 1 WU per GPU policy. Your work will get done faster and more concurrent GPUs will be working at once. A win win for everyone.

A good idea. They have lacked the means (or will) to limit the WUs to only the fastest GPUs, so limiting the number would help a lot.

By the way, if you can make the tile more descriptive, such as:
"When the supply of work units is short, limit the number to each cruncher", then that would attract more attention. It is a little vague now.