GPUHammer - Rowhammer Attacks on GPU Memories are Practical

Explore the first successful Rowhammer attack targeting NVIDIA GPUs with GDDR6 memory in this 14-minute conference presentation from USENIX Security '25. Learn how researchers from the University of Toronto developed GPUHammer, a groundbreaking attack that overcomes unique challenges in GPU memory systems including proprietary GDDR bank and row mappings, high memory latency, faster refresh rates, and built-in mitigations. Discover the novel techniques used to reverse-engineer GDDR DRAM row mappings and GPU-specific memory access optimizations that amplify hammering intensity to bypass existing protections. Understand how the attack successfully injects up to 8 bit-flips across 4 DRAM banks on an NVIDIA A6000 GPU and examine the practical implications for machine learning security, including demonstrations of how attackers can manipulate ML models to cause accuracy drops of up to 80%. Gain insights into this previously unexplored vulnerability in GPU memories that are critical for emerging machine learning applications and the security implications for systems relying on discrete GPUs.