Towards spatial fault resilience in array processors

Computing with large die-size graphical processors (that need huge arrays of identical structures) in the late CMOS era is abounding with challenges due to spatial non-idealities arising from chip-to-chip and within-chip variation of MOSFET threshold voltage. In this paper, we propose a machine learning based software-framework for in-situ prediction and correction of computation corrupted due to threshold voltage variation of transistors. Based on semi-supervised training imparted to a fully connected cascade feed-forward neural network (FCCFF-NN), the NN makes an accurate prediction of the underlying hardware, creating a spatial map of faulty processing elements (PE). The faulty elements identified by the NN are avoided in future computing. Further, any transient faults occurring over and above these spatial faults are tracked, and corrected if the number of PEs involved in a particle strike is above a preset threshold. For the purposes of experimental validation, we consider a 256 × 256 array of PE. Each PE is comprised of a multiply-accumulate (MAC) block with three 8 bit registers (two for inputs and one for storing the computed result). One thousand instances of this processor array are created and PEs in each instance are randomly perturbed with threshold voltage variation. Common image processing operations such as low pass filtering and edge enhancement are performed on each of these 1000 instances. A fraction of these images (about 10%) is used to train the NN for spatial non-idealities. Based on this training, the NN is able to accurately predict the spatial extremities in 95% of all the remaining 90% of the cases. The proposed NN based error tolerance results in superior quality images whose degradation is no longer visually perceptible.