@kcpeppe It's probably the first PiM proc they know of. Don't estimate the degree of amnesia in our field.
Another one: found a nice paper (good conf too) presenting the marvelous idea of hashed page tables. Almost the same (slightly different hash structure) as the PPC601 from 1991.
@phrygiandomina@johnmark_taylor I didn't write "LLMs are a dead end" (no mention of llms whatsoever). Just a quip about the limited computational primitives available due to path dependency.
e.g. people come up with various tricks to avoid Horner scheme on GPUs, but CGRAs can process them at full throughput.
@phrygiandomina@johnmark_taylor Not getting there. Consider KANs (2024). Claim substantial param count reduction. Let's assume they exaggerate and only cut param count in half.
They're still a dead-end since they're inefficient on GPUs. Due to the incestuous co-evolution of GPUs and NN, only matmul remains.
@kanzure Sarin was first synthesized with 1930s tech. Within capabilities of determined cult (1994).
Tech bros forget that not "everything is computer". You also need lab, and experience so you don't kill yourself prematurely.
@fentrocity@lougrims Shielding is basically more mass so massive increase of launch cost.
Radiation hardened chips exist, but have larger feature sizes. E.g. a space hardened riscv (not yet in prod) is a 12nm node, comparable to a 2020 CPU. This is btw a step up from the common rad hard PPC750.