[dalke]

I don't understand the question, as I don't know what "instruction" means in a portable way. AMD/Intel chips have instructions like LZCNT and CRC32, that don't exist as an instruction on R700-Family Instruction Set Architecture (nor other GPUs?).

Even if two functions do mostly the same thing (eg, multiply two floats), doesn't the Intel architecture have more complete support for the optional alternate exception handling of IEEE 754? If so, then they aren't really identical.

So, which instructions do you think are equivalent enough for your comparison?

Performance is driven by economics. Find where the economics between GPUs and CPUs are different, and you'll likely find where the performance inversion is.