Abstract: Due to the large memory capacity and frequent memory access, memory-hard hash functions are ASIC-resistance and may become the basis of the next generation password hash algorithms. Aiming at the password recovery problem of the memory-hard hash functions, this paper proposes two optimization methods for these functions based on the features of the GPU architecture. Firstly, two types of memory layout (interleaved and sequential) are analyzed and compared. Secondly, multiple threads are used for one hash function instance with the high-performance warp shuffling instructions on GPUs. Based on the sequential memory layout and corporation of four threads, we optimize two typical memory-hard hash functions (Scrypt and Argon2d) on the GPUs. The performance of Scrypt reaches 203 times the implementation of the hashcat software. At the same time, we also analyze the impact of memory capacity on performance, and point out that the reduction in the number of executable threads is the main reason for significant performance degradation on GPUs.

Key words: memory-hard hash functions, graphic process units(GPUs), password recovery