[cvoss]

But I think we are still far from approaching the stable post-Moore state. Being post-Moore is forcing engineers into exploring wackier and wackier architectures to squeeze out what computing power they can within the physical limits. Post-Moore life demands that we use parallelism anywhere and everywhere we can. A high degree of parallelism often requires a large memory set. Put those things together, and you're intensely battling data movement latency and cache consistency and contention issues. A solution that has recently emerged: rather than have each (hardware) thread tied to a core that has to manage the movement and consistency of data from RAM into nearby caches and back, instead have the threads migrate to the data. [0] Naturally, a lot of work is needed on the compiler side of language design to support this well without burdening the programmer. I don't foresee these kinds of new hardware ideas stabilizing any time soon.

[0] https://csmd.ornl.gov/highlight/designing-algorithms-emu-mig...