May 3, 2026

Created in May 03, 2026

2026   ·   gpu-performance   raytracing   gpu-security   rowhammer   ·   daily-log

Goals

  • Compare the performance of BoS prefetcher against baseline and cooprt.
  • Learn about the internal workings of OpenAI triton in order to develop an attack.

What I worked on

  • Compared the performance of BoS prefetcher with baseline and CoopRT.
  • Ran simulations for a configuration with both CoopRT and BoS prefetcher enabled to check if we are able to bypass CoopRT.
  • Setup Triton using the nvidia docker container toolkit, ran some prelimnary experiments

Key Insights

  • Today I learnt that triton has it’s own intermediate representation, similar to the IR emitted by LLVM.

Questions / Confusions

  • For the threat model of my attack I am considering scenario with a hostile user in an HPC clusters as HPC clusters do not enforce virtualization as strongly as the cloud providers.

Ideas / Hypotheses

  • For my next security work either I can attack the llvm backend or the software caches used by tools like triton.

Papers / Resources

  • None

Next Steps

  • If BoS prefetcher interacts positevely with CoopRT, I will try to incorporate RL in order to check if I can achieve better results than just naively combining both the ideas.

Introspection

  • Today was Sunday, so I was not expecting to be very productive, but alas, the day turned out to be insightful :)