The term “mainstream” in the context of the largest systems on the planet still only means a select few. But it is quite possible that proofs of concept for the sites building those huge future machines could set the stage for what’s next in some of the most demanding compute environments.
Right now, the options seem to be riding the decline of Moore’s Law or getting on the quantum computing train. The first is inevitable. The second is far from general purpose, especially for the largest government labs with decades-long code bases for nuclear and other mission-critical simulations that are tailored to traditional computing. But there is actually a middle route. While it has its challenges and its own scaling limitations, it could stretch supercomputing performance gains out a while longer while the world (ostensibly) gets ready for practical quantum or whatever is truly next.
Here’s why this middle option, an ultra-heterogeneous, customizable, waferscale platform, makes sense, and also why it didn’t until relatively recently, both from a technical and a market perspective: