Optimus Time

TICK TOCK, time is compute ⏱️ Baseline? Just the starting line 🏁

Comment

Inspirations

  • Optimizing to achieve higher speedups and lower runtimes
  • Providing support for systems that support CUDA and those that do not

What it does

  • Benchmarks the UserOptimizedTransformer against the Baseline Transformer, ensuring correctness within the given error limits and better performance with non-CUDA and CUDA support

How we built it

  • Non-CUDA implementation was done through manual prompting on ChatGPT web interface (Go subscription)
  • CUDA implementation was done through agentic capabilities of Codex (subscription provided by NUS)

Challenges we ran into

  • One of us didn't have access to an Nvidia GPU so we had to separate our development workflows then combine both solutions at the end
  • Understanding of the transformer model and its complexity within the given time frame
  • Sometimes Codex doesn't adhere to what was asked of it (e.g. asking for a CUDA implementation but it gives one that doesn't use CUDA)

Accomplishments that we're proud of

  • Providing a non-CUDA implementation and a CUDA implementation, both with substantial speedups
  • Being able to complete this in the short amount of time given and balancing school work at the same time

What we learned

  • Learnt more about CUDA and the various specific optimizations available
  • Learnt more about how to develop fast in a Linux environment
  • Learnt about Triton though we didn't use it

What's next for Optimus Time

  • HIGHER SPEEDUPS
  • The sky (AI subscription) is the limit

Built With

Share this project:

Updates

Submission history