Use of memristive memory arrays to achieve true Computation-in-Memory
The Memory Wall of computing stems from the fact that not enough data can be stored next to the processors and the ensuing data transfer delay stalls the processing. This can be seen in hiding memory latency with increasing on-chip cache sizes and levels and the number for concurrent threads. Moore’s law is not a solution, as the delay of on-chip wiring grows quadratically with scaling and as a result increasing amount of energy will have to be used to achieve sufficient bandwidth and delay. To compound the problem, neuromorphic and machine learning applications generate an enormous volume of data. Data are generated too rapidly to be processed within the energy, form-factor, and bandwidth limitations of current-day electronic systems. This project looks at bypassing the computational Memory Wall with the use of memristive, specifically Resistive RAM (RRAM), memory arrays thereby achieving true Computation-in-Memory (CiM).
Senior Research Fellow, Adjunct prof. of Nanoelectronics and Integrated Digital Systems Design