This is backward -- this sort of serialization is overwhelmingly bottlenecked on bandwidth (not CPU). (Multi-core) compute improvements have been outpacing band

This is backward -- this sort of serialization is overwhelmingly bottlenecked on bandwidth (not CPU). (Multi-core) compute improvements have been outpacing bandwidth improvements for decades and have not stopped. Serialization is a bottleneck because compute is fast/cheap and bandwidth is precious. This is also reflected in the relative energy to move bytes being increasingly larger than the energy to do some arithmetic on those bytes.