Agentic Kernel Generation for Diverse AI Accelerators

Agentic Kernel Generation for Diverse AI Accelerators

Abstract-
Production inference increasingly targets a heterogeneous mix of accelerators. Agentic pipelines interleave reasoning, tool calls, and multi-agent coordination, each with distinct compute and memory profiles. For optimal efficiency, each stage should run on the accelerator best suited to it. This shifts the bottleneck to infrastructure, because every pipeline now requires high-performance kernels across a growing set of hardware backends and programming models. Writing these by hand is time-consuming, demands deep low-level expertise, and does not scale as kernel complexity grows. This motivates program synthesis for diverse AI hardware. Large language models have shown strong code generation capabilities, but low-level performance code and cross-backend generalization remain challenging. I will present KForge, an agentic framework for automated kernel generation, show case studies across backends, and conclude with future directions.

Bio of Ankita Nayak:
Ankita Nayak is a Member of Technical Staff at Gimlet Labs, where she drives efforts on efficient systems for agentic AI and agentic AI for systems. Previously she was a Principal Researcher at Qualcomm AI Research, leading research in on-device generative AI and ML for AI-enabled 5G modems. She has published at leading ML, systems, and architecture conferences, and holds a Ph.D. in Electrical Engineering from Stanford University.