A Blueprint for Machine Learning Accelerators Using Silicon Dangling Bonds

As we approach the limit of transistor scaling, an appealing alternative in the form of quantum dots made of silicon dangling bonds (SiDBs) has been experimentally demonstrated to be capable of realizing sub-30 nm
²
logic gates. The introduction of SiQAD, a calibrated computer-aided design tool for the design and simulation of SiDBs, has further enabled the rapid exploration of this novel design space outside of experimental laboratories. Motivated by these advances and by identifying recent demands in machine learning acceleration, this paper proposes an architecture for an SiDB inference accelerator. Area and power estimates are made based on existing logic components and power models, the results are compared against Google's TPUv1. At the same clock rate, the proposed SiDB inference accelerator offers up to 10× improvement in area efficiency and orders of magnitude improvement in power efficiency, showing tremendous promise for further research into this novel platform technology.