I'm not actually sure who coined the term 'Race to Sleep' but I seem to hear it more often these days.
The idea behind it is to speed up an operation to be able to go into a very low power sleep state quicker after the operation at the expense of a higher peak power requirement during the operation itself. When 'Race to Sleep' works the overall energy required for the the faster execution + longer sleep time (as a reward) is lower compared to a previous architecture in which the operation took longer with less peak power drawn but a shorter sleep time. The 'operation' can be just about anything: Raw computing power, more complexity to speed up data transmission, GPU power, etc.
Does this really work in practice or is it just a myth? It seems it can work and AnandTech wrote a very detailed post on this phenomenon comparing power consumption for the same operations between a mobile device and its successor version. Have a look here for the details.
But at the end of the post he also notes that in practice, the gain when for example downloading and rendering a web page faster with higher power requirements and then make up for it by being in a sleep state for a longer time than before may be eaten quickly by users browsing the web more quickly because pages are loaded more quickly and thus they can start scrolling earlier.
So perhaps 'Race to Sleep' is most effective when a task that is sped up does not result in extra power expenditure later on due to the user being able to interact with a device even more quickly than before.