GPUHammer Attack Severely Impairs AI Accuracy on Nvidia GDDR6 GPUs
Researchers from the University of Toronto demonstrated that GPUHammer, a RowHammer-style attack, can induce single-bit flips in GDDR6 memory on NVIDIA RTX A6000 GPUs, drastically reducing AI model accuracy from 80% to below 1%. The attack exploits vulnerabilities in DRAM banks, but enabling ECC can mitigate the risk. The vulnerability affects certain high-end GPUs but not others like the RTX 5090 or data center cards with HBM memory.