So is there any reason why I would not want to run a separate instance of Claymore for each GPU? I'm using the "-di " switch to only select one card per instance.

I'm finding this much better when it comes to tuning a setup since one card crashing doesn't cause all cards to reload and waste precious mining time.

The only downside I'm seeing is reported hashrate on the pool is wrong since all instances have the same miner name but the calculated average looks spot on. Of course it might use a bit more memory and CPU but I don't think it's anything significantly important.