GPUHammer: Rowhammer Attacks on GPU Memories are Practical
summary
GPUHammer reverse‑engineers GDDR6 bank/row mapping, uses multi‑warp access patterns to reach 620k activations per refresh window, and aligns hammering to bypass TRR trackers. On an NVIDIA A6000 it injects 8 bit‑flips across 4 DRAM banks, identifies critical aggressor rows (most effective at ±2), measures thresholds of ~12–16k activations per victim row, and finds TRR eviction requires ≥17 aggressors; demonstrated ML model accuracy drops up to 80% and an exploit workflow that succeeds within ~10 attempts.
AI-written summary · about
chapters (10)
- 0:00Introduction
- 1:13Threat model for GPU Rowhammer
- 2:36GPU Rowhammer challenges
- 3:39Measuring memory layout
- 5:05Overcoming high latency
- 7:33Bypassing TRR defenses
- 10:01Observing bit flips
- 11:30Rowhammer thresholds
- 12:14TRR tracker implications
- 12:55Accuracy degradation attacks
related talks
40mRowhammer Attacks: A Walkthrough Guide
38mCSI:Rowhammer: Closing the Case of Half-Double and Beyond
29mCAn’t Touch This...
25mDRAMA: Exploiting DRAM Addressing for Cross-CPU Attacks
34mLeapFrog: The Rowhammer Instruction Skip Attack
25mOne Bit Flips, One Cloud Flops: Cross-VM Row Hammer Attacks...
18mSoftTRR: Protect Page Tables against Rowhammer Attacks using Software-only...
28mFlip Feng Shui: Hammering a Needle in the Software Stack