Sayani Majumdar1,2
VTT Technical Research Centre of Finland1,Tampere University2
Sayani Majumdar1,2
VTT Technical Research Centre of Finland1,Tampere University2
Real-time learning and adaptive ICT systems require computing beyond von Neumann realm. Neuromorphic computing, with human brain like energy efficiency and capacity of unstructured data handling emerged as one viable choice. High-performance neuromorphic computing hardware (HW) development requires exploration of a vast design space, for both suitable material combinations and their process optimization. Additionally, the sustainability of these materials, in terms of competitiveness in the electronic industry, compatibility with existing CMOS processes, dependence on critical supply chains, toxicity, recyclability etc. are difficult to predict from manual search of scientific literature and learning from repeated experimental cycles. Limited human resource and large experimental infrastructure costs, therefore, makes any new HW development extremely time-consuming and expensive. In this talk, I will present the idea of AI4AI, a data-driven approach to find low thermal budget and optimal semiconductors and ferroelectric materials for designing extremely energy-efficient non-volatile memories and synaptic weight elements. The AI4AI approach, first of its kind globally, where software AI is used to accelerate development of HW AI components, will not only expedite new HW development significantly, but will bring better material solutions to the ICT sector, considering sustainability and recyclability already at the design stage. While low thermal budget and low critical material dependence will ensure resource efficiency and green solutions for ICT HW, introduction of neuromorphic computing HW for AI and edge computing tasks will bring more than 1000-times energy efficiency compared to the current solutions.