Primary Investigator:
Aniket Kate

Project Members
Gustavo Franco Camilo, Ali Aqdas, Yanxue Jia, Aniket Kate, Mohammadkazem Taram

We present AES-GX, an architectural extension that brings native high-throughput AES to GPUs. Current GPUs, unlike CPUs, do not feature a dedicated AES instruction; thus, applications that rely on high-throughput AES need to either implement AES using existing GPU operations or rely on the CPU. Both are expensive. The former causes the AES instructions and the other application to compete for GPU resources such as memory and compute. For the latter, the data movement between CPU and GPU quickly becomes a bottleneck. AES-GX adds a single-fused AES instruction, designed through a comprehensive GPU-specific design space exploration that rethinks instruction granularity, key-schedule handling, and S-box implementation for the SIMT model. Our AES instruction is integrated as a Special Function Unit (SFU), leveraging existing operand collectors and scheduling to avoid intrusive pipeline changes while adding minimal area overhead of less than 1%. We implement AES-GX in a cycle-accurate GPU simulator and show that it delivers up to 58x speedups across security-critical GPU workloads, including state-of-the-art MPC protocols. Our results demonstrate that native AES support can significantly expand the class of cryptographic and privacy-preserving applications that can be efficiently accelerated on GPUs.