Optimise a VLIW SIMD kernel

Make an instruction-generating kernel faster on Anthropic's frozen simulator without changing its results.

Role: Software Engineer, Infrastructure
Difficulty: Hard
Practice time: Self-paced

The challenge

This exercise uses Anthropic's public performance take-home. You edit the kernel builder while the simulator, tests and interface remain fixed. Work toward a lower simulated cycle count, checking that each optimisation still produces the correct output. The challenge is self-paced and focuses on the generated program's execution rather than Python runtime speed.

What you'll practise

Instruction scheduling
Reason about how the generated instructions use the simulated machine.
Correctness under optimisation
Preserve the reference output while changing how the work executes.
Performance measurement
Compare cycle results rather than relying on how an optimisation looks in source code.

Your coding workspacePreview

perf_takehome.pyKernel implementation

# Your kernel implementation

# Check correctness and compare cycles

# Full starter code included

Benchmark results

Submit a correct kernel to measure its cycle count.

The kernel workspace, simulator and benchmark checks are included with Full Access.

Optimise a VLIW SIMD kernel

The challenge

What you'll practise

Instruction scheduling

Correctness under optimisation

Performance measurement

Investigate lost training throughput

Sort integers with bounded worker threads

Implement an LRU cache

The challenge

What you'll practise

Instruction scheduling

Correctness under optimisation

Performance measurement

Related challenges

Investigate lost training throughput

Sort integers with bounded worker threads

Implement an LRU cache