90/30 Club (ML reading) #39: Agentic Operator Generation for ML ASICs
Week #39: Agentic Operator Generation for ML ASICs
Join us at Mox to discuss Agentic Operator Generation for ML ASICs, a new paper introducing TritorX, an agentic AI system designed to automatically generate hardware-specific tensor operators for emerging machine learning accelerators.
This work presents a high-throughput kernel generation pipeline that uses large language models inside a structured feedback loop to generate Triton kernels, compile them, execute correctness tests, and iteratively debug failures. By relying on execution feedback rather than static prompting, TritorX can translate PyTorch operator specifications directly into production-ready accelerator kernels while ensuring correctness across tensor shapes, data types, and operator behaviors.
We’ll explore how a coverage-first design allows TritorX to scale backend development, successfully generating 481 PyTorch ATen operators and passing over 20,000 correctness tests, enabling large-scale models to run on Meta’s MTIA accelerator. We’ll also discuss how agentic workflows transform kernel engineering from a manual process into an automated, test-driven search problem, and what this means for hardware-software co-design and the future pace of AI infrastructure development.
🔎 Analyzed Paper
Agentic Operator Generation for ML ASICs
🔎Analyzed Papers
✍Google Drive for sharing Comments
Discussion at 20:00, (optional) quiet reading from 19:00.