Well, it's not that simple. The CUDA model has a inherent penalty for divergent thread behavior, that is, when all threads do the same, peak performance is att

Well, it's not that simple. The CUDA model has a inherent penalty for divergent thread behavior, that is, when all threads do the same, peak performance is attained. If they all follow a different path, the result can be disastrous. The relevance to media encoding is that the higher profiles do a lot of data dependent special casing. It's no accident that the earliest CUDA Video codecs only handled the base profile. This also hurts a CPU implementation but much less.