Any GPU miner crash after start with nFactor 19. Last fully stable is 15.
You seem to be right- at least for current nFactor:
I mostly get Hardware Errors on my GPUs. Rarely there is a accepted share, but with hashrates of 3Hash/s (on R7 240 4GB GPU), which equates performance of one modern CPU core.
Would it be possible to modify OpenCL code so that it would verify block hashes in parallel?
I presume if a GPU has 6 compute units and 2GB of memory it should be possible to do at least 6 parallel scrypt-jane block hashes (while one hash consumes 128MB of memory). The goal would be to offload CPU with chain validation.
I tried to dig into it, but it's a big bite for me.