Performance Evaluation of Kuznyechik and Magma Symmetric Ciphers on AMD and NVIDIA GPUs

In this paper we study the efficiency of GPU-based implementations of Kuznyechik and Magma symmetric ciphers written in CUDA and a new cross-platform framework called HIP. We tested our implementations using AMD Radeon RX Vega 56, NVIDIA GeForce GTX 1080Ti and NVIDIA GeForce RTX 2080Ti. We measured the clear encryption speed and the actual encryption speed on three different data transfer modes: an explicit copy mode, a zero-copy memory mode and copy overlapping using multiple command streams. Achieved results show that AMD Radeon RX Vega 56 outperforms both NVIDIA graphics processing units and delivers 178 Gbps of a clear encryption speed for Kuznyechik and 940 Gbps for Magma. A zero-copy memory mode is more efficient than explicit copy mode and allows to achieve the encryption speed around 70-75 Gbps for both ciphers on all used graphics processing units. Overlapping an encryption and a data transfer using 32 command streams allows to achieve the encryption speed around 79 Gbps for both ciphers on all used devices.