Error Recovery Through Partial Value Similarity

Abstract

Soft errors arose as a critical problem for microprocessor designers due to shrinking feature sizes and increasing clock rates. Many attempts appeared to protect the register file which is the main storage component in modern microprocessors. Exploiting existing replica values in the register file is a recently proposed method to correct errors that occur on the register values. In this paper, with the observation that partial matches are clustered in the least significant bytes of the similar values in the register file, we propose to extend the coverage of existing replica based error correction schemes. Our schemes almost double the coverage of previously proposed mechanisms. The proposed method can recover 39.8% of the stored values in the register file by using this similarity and single-bit parity protection. The power overhead of this method constitutes only 3.1% of the register file's power budget.