Shield: Cost-Effective Soft-Error Protection for Register Files

Device scaling and large integration increase the vulnerability of microprocessors to transient errors. One of the structures where errors can be most harmful is the register file, because it is part of the architectural state. Moreover, because the register file is frequently read, an error can quickly propagate to other parts of the processor. This paper presents Shield, a novel, cost-effective architecture that increases the resistance of register files to soft errors. Shield is based on two key observations. First, the data stored in a register is only useful for a small fraction of the register’s lifetime. Second, not all registers are equally vulnerable to soft errors — a small fraction of long-lived registers is more liable to errors. Shield selectively protects registers by storing the ECCs of the most vulnerable registers while they contain useful data, and checks their integrity off the critical path. Our experiments show that Shield reduces the architectural vulnerability factor of the integer register file by up to 84% (73% on average), and by up to 100% (85% on average) for the floating-point register file while affecting performance negligibly.