Not all cards finish in under 2 minutes. If the application would checkpoint every 2 minutes, or even every 1 minute, that'd be wonderful.

Tried it and it is not working with tthrottle, since it does not have "exclude gpu app" option.
Tip of the Hat to KSMooney, it would be nice if the application would checkpoint at 2 or 4 minutes, so the faster cards would not be bother with this, while helping the slower cards.

I do not think it is hard to implement something that it is already done in other projects, and it doesn't require to reinvent the wheel.
If this is too cumbersome to do, then I suggest that future apps should run one WU on each Compute Unit, so that every card should run them in "almost" the same amount of time.
If the fastest cards can run one task in under 2 minutes, than wrap 16 of them, and do a chekpoint every 2 minutes so that everybody is happy.
P.S. If memory serves me right, back when there were some hints about GPUs being used in WCG, you were saying that it cannot be done so easily or the speedup will not be as significant.

There's checkpointing requested on the single WU per GPU task, which is what I'm talking about [think my post was clear on that]. If the plan, yes there was a plan, and multiple were packaged in a GPU job, then checkpointing becomes sensible [too noted in my post]. What I said few years ago and what I'm saying now? Was talking as said a single HCC WU task and some members having a lag issue where checkpointing at arbitrary points is likely to worsen the situation. And YES, it depends on which WCG project is brought to the GPU. The HCC is integer intense [none/little FPO dealt with I guess in the CPU phases], integer is easy peasy for a GPU card, yet to get it to production took how long and what was all needed to be put in place? Some research done by scientists have shown the case that little to none will be gained by porting other sciences to the GPU [search posts on the forums]. If there was, and there would be enough time to make a "bring it to the grid" viable, it would have been done. For Rice a GPU development was done, as post-processing project and it was not deemed big enough or too involved to bring it to WCG, so the scientists are doing it in-house. The future... anything GPU able in this type of research is substantially to be moved off grid is one expectation. Much easier to have a homogenous set of [latest] hardware and let it churn through in a few months than trying to cater for a zillion different volunteer requirements and afterwards having to deal with just enough statistical variance that your set of results is just not optimal. This is the most major concern with public, distributed computing.