GDDRHammer: Greatly Disturbing DRAM Rows - Cross-Component Rowhammer Attacks From Modern GPUs

While Rowhammer has been extensively studied in CPU-based memory systems, a very recent work by Lin et al. (USENIX Security '25) extended this line of research to GDDR6 GPU memory, demonstrating the first Rowhammer bit flips on NVIDIA GPUs. However, they tested only a single GPU and observed just 8 flips across 4 DRAM banks. Moreover, their proof-of-concept exploit only demonstrated degradation of a deep neural network's inference, leaving both the extent of GPUs' susceptibility to Rowhammer and the impact of said bit flips largely unexplored. We address this gap by exploring both the prevalence and impact of Rowhammer on GPUs. First, we develop techniques for dramatically amplifying Rowhammer on modern GDDR6-based GPUs. By utilizing the inherent parallelism in GPUs and developing new techniques for bypassing Rowhammer mitigations on GPUs, we are able to produce 129 bit flips per DRAM bank on average, demonstrating a 64x increase over prior work. We also present a comprehensive characterization of GPUs' Rowhammer susceptibility by testing more than 25 GPUs across multiple systems. Our findings show that nearly all tested RTX A6000 GPUs remain vulnerable under realistic configurations despite hardware-level mitigations, demonstrating that Rowhammer is far more prevalent on GDDR6 memory than previously understood. We also show the impact of Rowhammer on GPUs by demonstrating the first GPU-to-CPU Rowhammer exploit, with a practical end-to-end attack wherein an attacker flips bits in the GPU's memory and gains read and write access to all of the host CPU's memory.