HARDWARE CHAT ⚙🏁

The irony at the center of analog AI chips: the multiply is free, but the ADCs converting the result back to digital often dominate the energy and area budget Analog in-memory computing is having a moment again (EnCharge with TSMC, Mythic's $125M round, IBM's HERMES chip), and the pitch is interesting: collapse the von Neumann round trip by storing weights as conductances and letting physics do the MAC.What gets left out of most coverage is that the analog/digital boundary is still there. Voltages in, currents out, they need conversion, and in practice a meaningful share of the chip budget goes to the converters rather than the crossbar array.The other problem is drift. An analog weight is a material property, so a PCM or RRAM cell reads differently a week after it was written. That means periodic recalibration or reprogramming is a permanent overhead digital memory doesn't pay.I wrote up the mechanism plus a numpy simulation showing how programming noise, read noise, and a 4-bit ADC stack up to >8% relative error on a single layer, before any compounding: https://towardsdatascience.com/analog-ai-is-back-can-it-survive-its-own-noise/Disclosure: my article on Towards Data Science. Via Reddit 💬 | @HW_HUB 🖥