ML systems · Duke CEI Lab / 2026
Fornax
Software for a diffusion accelerator
The problem
A custom accelerator needs software that translates model operations into work the hardware can execute. For diffusion workloads, this means accounting for instruction support, memory constraints, and numerical behavior together.
My contribution
At Duke’s CEI Lab, I developed C++/Python compiler tooling within an MLIR-based stack, collaborating with algorithm and hardware researchers. I implemented operator splitting and lowering to hardware instruction and memory constraints, and contributed to quantization-aware training and nonlinear function fitting.
What connects the work
This project brings model behavior and hardware constraints into the same engineering process. The software contribution spans how operations are represented, how they are divided into executable pieces, and how numerical approximations fit the target hardware.
Current status
Fornax was submitted to ISSCC 2027, with me as a co-author. This page describes my work at a high level; the project implementation is not linked publicly.
Want to talk about this work? Get in touch ↗