Helion tops B200 kernel hackathon

The result is a concrete proof point for Helion's core claim: serious GPU kernel performance is reachable without being a Triton or CUDA expert, just a PyTorch programmer who understands tiling.

• –Helion's autotuner systematically explores block sizes, loop orderings, and memory layouts — a search space that explodes combinatorially on B200 hardware
• –B200 kernel optimization is brutal: Gated DeltaNet patterns, Mixture of Experts, inter/intra-chunk attention, and KV caching each demand different strategies per model architecture
• –The last 5–10% still required manual grinding — Helion compresses the hard work dramatically but doesn't eliminate expertise entirely
• –Local inference via NVIDIA Pro 6000 powering an agent harness performed well throughout, reinforcing that local LLM setups are viable for competitive development workflows
• –Hackathon submission repo published at github.com/brandonin/helion-hackathon-submission, useful as a reference for anyone exploring Helion on convolution kernels