Yes, I agree with that. But it also takes a lot of software work. Getting decent performance on a GPU for an existing scientific application requires something

Yes, I agree with that. But it also takes a lot of software work. Getting decent performance on a GPU for an existing scientific application requires something closer to a full rewrite than the usual porting effort to a novel CPU architecture. That is one of the reasons that people are writing new scientific applications to target GPUs from the beginning rather than just adapting mature older programs to GPUs. (Though there is some of both approaches.)