Supply chain / In-memory and wafer-scale
IBM Research analog AI
An IBM research programme on analogue in-memory computing, where neural-network weights are held in memory devices that also do the arithmetic, cutting data movement.
- Type
- Technology
- Where it sits
- A long-range route to lower-energy inference; it signals where accelerator design may go rather than offering capacity now.
- Constraint
- The project page describes prototype chips and research results only; it lists no product or customer access.
- Maturity
- Research
What we checked
- IBM's inference work uses phase-change memory cells, and the page says its chips hold over 13 million of them. (source)
- The page records prototype chips in 2023 that used much less power for AI tasks and software-equivalent accuracy on transformer networks in 2021. (source)
Written by ComputeFinder from public pages on 2 October 2026. Company statements are the company's own, not independently checked, and reported items are marked as such. Not yet reviewed line by line by a second person. If something is wrong, tell us.