hn.today

Dissecting the GPU Load/Store Unit and Exposing Its Security Risks

arxiv.org4 points0 comments
Screenshot of Dissecting the GPU Load/Store Unit and Exposing Its Security Risks

This study reverse-engineers how load/store units (LSUs) handle memory instructions on recent NVIDIA GPUs, a part of the hardware that affects both kernel speed and security. Using microbenchmarks, the researchers derive a cycle-accurate model for global-memory L1 cache hits and shared-memory accesses. The model shows that access latency depends on addresses, particularly through shared-memory bank conflicts and L1 tag lookups, and that loads can contend at both points. They also find that when the LSU pipeline is saturated, a warp can prevent a warp from a later-launched block from making progress. These measurements fill gaps in understanding how loads are serviced and where workloads interfere.

The researchers use these mechanisms to demonstrate security and performance risks beyond previously known GPU attacks. Either global-memory bank conflicts or L1 tag lookups can, on their own, provide enough timing information to recover a complete AES-128 key under the threat model of earlier GPU timing attacks. Under NVIDIA’s Multi-Process Service, both mechanisms also enable covert communication between processes, with channel rates increasing across GPU generations. Finally, LSU contention can cause severe denial of service: a single warp slows a colocated cuBLAS matrix multiplication by as much as 750 times on an A100 and 563 times on an H100. The findings suggest that LSU behavior creates multiple exploitable channels, not merely a performance bottleneck.

Read on arxiv.org0 comments on Hacker News

Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.

More in Security

The daily digest

Today's best Hacker News stories, summarized and screenshotted, one email a day.