Early research from IBM Research Europe shows a new phase-change memory design that stores multiple AI data vectors superposed in the same analog hardware. Prototype chips ran neural network inference with accuracy close to full-precision software, while letting engineers tune how much accuracy to trade for extra storage.
Running large AI models is expensive largely because moving data between memory and processors costs time and energy. Packing more vectors into the same physical memory could let AI systems run on smaller, lower-power hardware — from data centers to phones — without a big accuracy penalty. The tunable tradeoff means designers can pick the balance for each use.
✦ Storing more in less hardware could make AI cheaper and more efficient.