Level 3 · Performance engineering with a public bar
Can you optimise the take-home kernel on its simulated machine below Anthropic’s published 1,487-cycle threshold?
Anthropic open-sourced its original performance-engineering take-home: optimise a kernel for a custom simulated VLIW SIMD machine. Below 1,487 cycles, you can email your code to Anthropic’s performance recruiting team.
It is a real hiring test with a published bar and a direct route to a recruiter. Anthropic says about 1,000 candidates completed it and it helped hire most of its performance engineering team. It needs no GPU at all.
01Clone the repo and read everything
Read the README, problem.py (the machine) and perf_takehome.py (the code you optimise). The starter code is deliberately slow – about 147,700 cycles.
$ git clone https://github.com/anthropics/original_performance_takehome.git02Run the tests to see where you stand
The tests print your cycle count and which published thresholds you pass.
$ python tests/submission_tests.py03Model the machine on paper
Write down the instruction set, how many of each operation can issue per cycle, vector width, and memory rules. Work out a theoretical lower bound for the kernel.
04Use the trace viewer
Find where cycles go and which execution units sit idle.
05Optimise in measured steps
Vectorise, then pack independent operations into the same cycle, then reduce memory traffic. Commit after each step with the new cycle count in the message.
06Know the published milestones
For the original 2-hour version, the README lists Claude scores: 1,790 (Opus 4.5, casual session, about the best human 2-hour result), 1,579, 1,548, 1,487 (Opus 4.5 after 11.5 hours) and 1,363 (improved harness). The best human result is ‘substantially better’.
07Verify you haven’t touched the tests
The README warns that LLMs can cheat. If you use an AI assistant, tell it not to modify tests/. This must print nothing:
$ git diff origin/main tests/08Submit if you beat 1,487
Email performance-recruiting@anthropic.com with your code and ideally your CV, and say you ran both checks. The threshold may change with new model releases – read the README first.
09Stretch: aim below 1,363
Don’t stop at 1,487. 1,363 is Claude Opus 4.5’s best after many hours in an improved harness, and others have published sub-1,300 results, so the bar is moving. Check the README first – Anthropic says the threshold may change.
10Write it up either way
Code plus a short write-up: trace screenshots, each optimisation step and its cycle gain, and what you tried that didn’t work.