Physics-Consistent Thermal SPICE and Multi-Correlated Recurrent Neural Networks to Simulate Sophisticated FinFET Circuitry
TimeTuesday, July 12th6pm - 7pm PDT
LocationLevel 2 Lobby
Event Type
Networking Reception
Work-in-Progress Poster
DescriptionA physics-based thermal SPICE model composed of distributed thermal R-C network is used to simulate the self-heating of FinFET folded inverter chains up to 37 stages. For circuits with more stages, a series of multi-correlated recurrent neural networks (RNNs) is used to predict the temperature profile. The correlated RNNs trained by the SPICE data with stage number (stage#) ≤ 17 can predict circuits up to 37 stages (2.2X SPICE) with the error as low as 0.9oC, indicating that the thermal physics is learned by NNs. NNs can thus predict 82 stages confidently, while SPICE cannot simulate more than 37 stages.