Event ·Official first-hand

LLVMhammer:通过 Rowhammer 位翻转让 LLM 推理变慢 4-16%

研究者提出 LLVMhammer,这是首个利用 Rowhammer 位翻转攻击拖慢 LLM 推理的方法,攻击目标是 Triton JIT 编译管线的优化 pass。位翻转让 Triton 生成功能相同但性能更差的 GPU kernel,且由于 Triton 会把编译好的 kernel 缓存到磁盘,一次攻击就能造成持久性变慢。实验在 NVIDIA RTX 4090 和 AMD RX 7900 XTX 上测试了 Qwen2.5-7B-Instruct、Llama-3.1-8B-Instruct、Mistral-7B-Instruct-v0.3、Deepseek-llm-7B-chat 四个模型,单 GPU 推理延迟增加 4-16%。位翻转还能破坏输出质量,BERTScore 最低降到 25%。

Current read

有人用 Rowhammer 硬件攻击搞坏了 Triton 编译,四个主流模型实测慢 4-16%,搞 GPU 推理的都该看看这篇。

11 sources36° AI heatLast updated 2026/10/22 06:24:35

Evidence

11 sources
Primary sourcearXiv: DeepSeek
View source
Related source 1kimmonismus
View source
Related source 2Teortaxes
View source
Related source 4@atomic_chat_hq
View source
Related source 5The Information
View source
Related source 6marktechpost
View source
Related source 7Yuchen Jin
View source
Related source 8Nathan Lambert
View source
Related source 10rohanpaul_ai
View source

Conflict check

Dedup data only shows that these sources discuss the same event, not that they agree. There is no structured supporting or refuting evidence yet, so no conflict is inferred automatically.