GPUHammer: Practical Rowhammer Attacks on GPU Memory
Researchers from the University of Toronto have developed GPUHammer, the first successful Rowhammer attack against GPU memory, specifically GDDR6 memory in an NVIDIA A6000 GPU. The attack uses user-level CUDA code to bypass in-DRAM defenses like TRR, inducing bit flips across all tested DRAM banks. This allows a malicious user to tamper with other users' data in shared, time-sliced environments. A proof-of-concept demonstrated an accuracy degradation attack against a victim's DNN model, reducing accuracy from 80% to 0.1% with a single bit flip. While enabling ECC mitigates the risk, it can introduce up to a 10% slowdown for ML inference workloads on the A6000.
Read more