With ethash, about half of the GPU compute resources are "wasted" waiting for memory accesses. With AMD GPUs, it is possible to have the kernel code load data from RAM into a register and continue to execute other instructions that don't use the data from RAM. A dual mining kernel takes advantage of that by a different algo during the dead time waiting for the RAM latency.
Thanks for your explanation, that's interesting. So this dual mining concept is specific to the core functionality of how EthHash works.
Interesting, just wondering if you put the numbers for Decred too, would be nice to know.