Speaker
Description
We propose a fault tolerant mechanism for linear DSP applications, that relies on: (i) parity based real number error correction codes to store redundant information efficiently, (2) gradient descent correction loops for error estimation, and (3) fine grained check-pointing and roll-back for consistent redundant information encoding. The error correction is performed by a novel gradient descent symbol update method designed for real number parity based codes, which performs error extraction from the syndrome vector.
We use the linear transform property of the application in order to compute the correction offset that is applied to the output data. The proposed solution is validated for a 512 point memory-based FFT architecture, augmented by the means of an LDPC code, for which an efficient way to embed the processing associated to the error correction code - encoding, syndrome check, decoding – is devised.
