Speaker
Description
GPUs were initially developed to facilitate the rendering of graphic elements, but as applications of the SIMD model appeared they came to be used in increasingly diverse fields, and more recently in the rapid development of AI. Frameworks such as CUDA expose many capabilities, but technical details regarding memory organization and protection mechanisms are often missing from the official publications, which is concerning given the increasing sensitivity of the data processed on GPUs. This paper investigates the memory organization and the vulnerabilities present on the NVIDIA Turing architecture, using a GeForce GTX 1650 Mobile with compute capability 7.5, and verifies whether vulnerabilities reported on previous generations have been addressed. We show that there is no protection for any overflow in device memory: if the memory is paged, it can be accessed, and the heap is only an abstraction over global memory with no extra security measures. We prove that the instruction code of a dynamically loaded module can be obtained from a normal kernel. Local memory can be overflowed in both directions, and threads can read values that were not properly zero-ed by threads from previous blocks, from which we deduce a mapping stride of 32. We also show that the dynamic stack and the stack holding call frames are the same region, and that the saved instruction pointer resides on it. Finally we quantify bank conflicts, measure L1, L2 and DRAM access times, determine the L2 cache geometry and recover the lower bits of the XOR function that decides the set index.