A new technical paper from the University of Toronto describes GPUThor, a Rowhammer attack designed for NVIDIA GPUs. The method relies on non-uniform hammering patterns, which the researchers say can generate higher-intensity memory disturbance than conventional uniform approaches.
The attack specifically targets GPUs with error-correcting code (ECC) memory, a feature that normally detects and corrects bit flips. By amplifying the hammering effect, GPUThor can overcome that protection, suggesting that ECC alone is not a sufficient defense against Rowhammer on modern accelerators.
The findings are significant as GPUs are increasingly shared across cloud tenants and used for AI training, where memory corruption could lead to data leaks or privilege escalation. The paper is a research result, so practical exploitation would depend on system configuration, but it underscores the need for stronger memory hardening in accelerator hardware.