Learning to Formulate: Agentic Hardware-Software Co-Design of AI Infrastructure through Progressive Optimization and Verification
Formulaic lets an AI agent define the design problem, while a solver, verification and measurement test the resulting choices.
The agent progressively refines mathematical formulations from the computation graph toward the chip, using program templates with explicit correctness conditions. Agent-written templates ran complete Qwen3-8B layers 1.075× faster than SGLang on NVIDIA GB10. In a simulated Ventus GPU study, joint hardware-software design reduced worst-scenario cycles to 42.8% of the reference, compared with 93.4% for software optimization alone. Reusing proofs made checking a five-step GEMM refinement chain 2.51× faster; these were separate studies, not an end-to-end demonstration.
ArXiv · AI/CL/LG's note
The agent progressively refines mathematical formulations from the computation graph toward the chip, using program templates with explicit correctness conditions. Agent-written templates ran complete Qwen3-8B layers 1.075× faster than SGLang on NVIDIA GB10. In a simulated Ventus GPU study, joint hardware-software design reduced worst-scenario cycles to 42.8% of the reference, compared with 93.4% for software optimization alone. Reusing proofs made checking a five-step GEMM refinement chain 2.51× faster; these were separate studies, not an end-to-end demonstration.
ArXiv · AI/CL/LG's note
score 5