It can be worth revisiting these older (1990s) ideas for large-scale parallelism. They mostly died on the vine at the time, because most people only had 2 or 4

It can be worth revisiting these older (1990s) ideas for large-scale parallelism. They mostly died on the vine at the time, because most people only had 2 or 4 cores to play with. If you could afford a CM-5, you might get 1000 cores (but shared among your whole institution.) Now it's possible to play with a million GPU cores with a modest budget. The main constraint is figuring out how to use them.