What if it would be more efficient to make each miner to try to find nonces for different sets of transactions rather than finding nonces for the same set?
The nonce is iterated to find a block hash lower than the target.
Before you start trying nonces you must have a candidate block already (complete with transactions)
That's right, but I heard in a lecture on Bitcoin mining that in case a miner can't guess the needed nonce for too long it includes more transactions in the block(or excludes some of them) to change the Merkle root and to start guessing nonces for the new one. Is that info correct? And if yes then that's what I was asking about.
Also the reply by DannyHamilton is pretty much about the same thing:
when miner divide the work to each gpu,
does it divide in a way that they won't overlap?
In most cases, I would expect that the mining software would simply build completely different blocks for each, rather than trying to split up nonce range.
And one more thing. I think that there are so many variants of possible nonces that being picked randomly they almost never overlap.