• 1 Post
  • 646 Comments
Joined 1 year ago
cake
Cake day: June 5th, 2025

help-circle







  • Look at the enormous processing resources of biological brains. Human brain is 2% of body mass but 20% of baseline metabolism – this is very expensive evolutionary. Neural hardware used for LLMs or just any scientific numerics accelerator is just a bad reinvention. Your argument reminds me of Minsky’s “5 MIPS is enough for AI”. Nope. You have to track a lot of state, its relationships and refresh it all very quickly. Computation is expensive.





  • The node sizes are backwards computed from the unit area of a one bit SRAM cell. It used to be the physical size of the smallest structure, e.g. a transistor gate.

    I agree it’s highly misleading. Also, people forget that the original Moore’s law publication mention constant doubling time of transistors per area per unit of cost. So people keep pointing at stacking and 5 nm WSI like Cerebras to argue Moore scaling didn’t die a long time, but ignore costs.