Probing Vulnerabilities in GPU Shader Execution

Graphics processing units (GPUs) are increasingly used for general-purpose parallel computing. However, the rise of GPUs in security-sensitive domains has exposed vulnerabilities arising from architectural weaknesses. In particular, side-channel attacks on shader execution can leak sensitive data through mechanisms like timing analysis, contention monitoring, and access pattern observation. We present a detailed investigation into shader-based side channels on commercial GPUs. We introduce a taxonomy of attack vectors categorised as timing, contention, and access pattern channels. We use standard shader interfaces like CUDA and OpenCL to demonstrate proof-of-concept exploits on NVIDIA and AMD GPUs for each attack class. Our analysis reveals pervasive side-channel risks from fundamental GPU architectural traits like massive multithreading, memory hierarchies, and throughput-optimised pipelines. Results show that even academically proposed countermeasures are inadequate in practice. We conclude with an in-depth discussion on the root causes of GPU shader vulnerabilities to inform specialised hardware-software co-designed defences. We comprehensively evaluate real-world GPU shader execution risks to guide further research on security-aware accelerator design.