I have never seen that error before.
It's got nothing to do with the GPU's precision. (Actually, the 580 is *the* optimal card for running GeneferCUDA -- it is the fastest device in the world for computing GFNs).
I looked up that error in the BOINC client source code and it happens when, during the round-robin work simulation, the number of available FLOPS on your computer goes negative.
To be honest, I've never paid too much attention to how the internal scheduling in the BOINC client works, but it looks like the client has detected that you have too much work to complete before the deadlines expire.
Given that you have 10 WR tasks and 2 short tasks, I'd have to agree with the scheduler. Even on a 580, those WR tasks will take about 5 days each. So with a pair of 580s, it's probably not possible to finish 10 in time. 8, yes, but 10 will be really, really close.
You might want to try lowering your cache size and aborting a pair of those WUs, at least until BOINC gets a better handle on how long those WUs really take. (If you're running other PrimeGrid sub-projects, that gets convoluted because BOINC doesn't comprehend that one project can have WUs with differing DCFs).
In any event, it looks like nothing bad actually happens because of that error, other than perhaps the scheduler not working the way it's intended. Of course, if you've been using BOINC for any amount of time, you're likely already used to the scheduler not working the way one would expect it to work.
My lucky number is 75898524288+1