Details
- Affiliation
-
AffiliationUniversity of Electronic Science and Technology of China
- Country
In edge computing, the requirement for power consumption are extremely harsh, such as biomedical detecting and wearable devices. However, traditional neural network processor contains many weight shift operations due to large number of parameters. Recently, near memory computing (NMC) can greatly reduce large data movements compared to the traditional Von Neumann architecture. Unfortunately, the NMC method cannot optimize the calculation accuracy and power consumption at the same time. To meet the need of ultra-low power consumption in biomedical detecting, several time-domain (TD) based computing engines have been developed to move calculation into both digital and time domain. In this work, we present a TD-based NMC biomedical processor, namely TDPRO, to optimize the detection of arrhythmia with extremely low power consumption. In our TDPRO, weight parameters is located inside processor without data movements. We design a computing engine to support highprecision multiplication and addition (MAC) with 8-bit input and 8-bit weight parameters.