NVIDIA GPU load/store units leak AES keys and stall colocated jobs
Original title: Dissecting the GPU Load/Store Unit and Exposing Its Security Risks
Bank conflicts or tag lookups alone recover a full AES-128 key, and covert-channel rates rise across GPU generations.
Researchers reverse-engineered how recent NVIDIA GPU load/store units serve loads, using microbenchmarks. They give a cycle-accurate latency formula and show loads contend on bank conflicts and L1 tag lookups. Through either channel they recover a full AES-128 key under a prior GPU AES timing-attack model. Under Multi-Process Service, both channels form cross-process covert channels, and pipeline starvation slowed colocated cuBLAS multiplication by up to 750x on the A100 and 563x on the H100.
Unconfirmed: Slowdown figures and channel rates are the authors' measurements, not independent reports.
- A full LSU pipeline can let one warp starve a warp from a later-launched block.
- Covert channels use bank access or tag lookup across processes under MPS.
- The reported cuBLAS slowdown is up to 750x on A100 and 563x on H100.
- The authors say prior attacks covered only isolated LSU mechanisms.
Impact: AES-128 key recovery, covert channels, and colocated workload slowdown
| Product | Affected |
|---|---|
| NVIDIA GPU load/store unit | recent NVIDIA GPUs, including A100 and H100 |
- Vendors
- NVIDIA
From the original article
1 min read at arxiv.orgGPUs are the main platform for large-scale machine learning and power the largest supercomputers. Their load/store units (LSUs) often determine kernel performance and have been exploited in prior attacks. Yet what is known about the LSU is only a few isolated mechanisms and parameters, and even a single one has been enough for an attack.
By Wenbo Han · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.