Machine learning on hardware: the SoC-Builder RTL dataset
A reproducible, machine-generated corpus of synthesizable SoC RTL — the data foundation for representation learning on hardware, and a step toward differentiable EDA.
Machine learning has transformed domains that had large, clean, labeled datasets. Hardware design mostly hasn't — not because the problems are uninteresting, but because the data is scarce, proprietary, and hard to reproduce. The SoC-Builder RTL dataset is my attempt to change that starting point.
Why generate the corpus
Real-world RTL is locked inside companies and comes with no consistent labels. To learn anything general about hardware, you need a corpus that is (1) large, (2) synthesizable, and (3) reproducible — so results can be checked and extended rather than taken on faith. So the dataset is machine-generated: a controlled process emits synthesizable System-on-Chip RTL designs, which means the labels are known by construction and the whole corpus can be regenerated.
What it's for, today and later
The v1 experiment release targets RTL representation learning — teaching models the structure of hardware designs the way code models learned the structure of software. As the corpus grows, the same foundation extends toward the harder, more valuable targets:
- netlist prediction — from RTL toward gates,
- timing prediction — estimating closure before you run the full flow,
- placement prediction — spatial structure without a full P&R run.
The bigger bet: differentiable EDA
This dataset is one piece of a longer thesis I keep coming back to: that the gap between high-level design intent and the physical implementation flow can be made differentiable — that models can learn to predict, and eventually guide, steps that today require slow, iterative tool runs. You can't build differentiable electronic design automation without data to ground it, and hardware doesn't have a GitHub-scale corpus lying around. So the first move is to build one.
It's an experiment release, deliberately. The point isn't a finished model — it's a reproducible foundation that other people can train on, poke holes in, and grow.