Lifetime Improvement Method for Memristor-Based Hyperdimensional Computing Accelerator

This study proposes a lifetime extension method for a hyperdimensional computing (HDC) inference accelerator implemented with memristors using error detection and self-repair techniques. The proposed method detects faulty memristors by calculating the sum of the memory values of memristors that store hypervectors in the HDC as a checksum and replacing the failed memristors with redundant memristors, thereby extending the lifetime of the HDC inference. When a memristor fails, the resistance is stacked at the minimum or maximum. This causes the checksum value to change when a failure occurs, which can be detected by comparison with the correct summed value. A numerical evaluation with a language-comparison task shows that the lifetime can be extended by five times while maintaining the same accuracy as in the case where no lifetime extension method is applied.