I haven't decided yet. The search up to 10^20 will take more than 4 years at current speed, so 10^21 will take more than 40 years which is just too much. So it all depends on how many volunteers will be running the project in 4 years from now.

On my ATI R9 2960X : runtime between 109 seconds and 1266 seconds for a task.
On my GTX 680 : runtime between 383 seconds and 2116 seconds for a task.

This is normal. All WUs were normalized for CPU version, not for GPU. Some WUs get a huge speed-up when running on GPU, some don't. It depends on how well a work unit can be parallelized.