May 4, 2026

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 + Cooprt with baseline and CoopRT for a resolution of 128x128.
  • Understood the compilation path of Triton

Key Insights

  • BoS prefetcher does not always interact positevely with CoopRT, this presents a unique oppurtunity (hehe) (muhaha)
  • Triton first gets the AST from python then compiles down to triton IR, then gets further compiled down to triton GPU specific IR, followed by LLVM IR, then PTX and finally SASS via ptxas.

Questions / Confusions

  • I need to look at the encoding used for all the intermediate representation so that I can decide attacking at which point would be the most lethal.

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

  • Figure out the root cause for negative interaction between CoopRT and BoS prefetcher

Introspection

  • Today was a sad day, two people got what they did not deserve.