Vector search is a great application for this. You can run the HNSW algorithm on the riscV cores local to the memory. Which is even better than GPU since it’s actually MIMD with tons of cores so each can follow its own branching logic. And a bunch of servers can share the RAM (read only) in the M1.