I Spent 1 Billion Tokens Making a Real Chip with AI — and Found Its True Limits
I Spent 1 Billion Tokens Making a Real Chip with AI — and Found Its True Limits TL;DR: This is not "I prompted an AI to make a chip." It's the consolidated retrospective of two month-long projects — a SHA-256 crypto accelerator driven end-to-end from RTL to a 10-item signoff on an open-source EDA stack, then integrated into a real SoC. Two AI agents plus one human, roughly 1 billion tokens across…
In this story, a researcher spent a total of one billion tokens over two months-long projects to create a SHA-256 crypto accelerator and integrate it into an SoC. The projects utilized an LLM agent, an executor AI, and a human to cover the entire design process, from RTL to signoff checks. While the AI agents were able to successfully write code, orchestrate flows, handle crashes, and perform multi-scale decompositions, there were limitations to their abilities.
The most significant issue was premature completion, where the AI agents accepted partial evidence as global conclusions. This often led to failures, such as undefined register states at power-up and issues with gate-level simulation. To mitigate this, the researcher implemented four rules: dual-agent independent review, crash and anomaly handling, multi-scale decomposition, and toolchain assumption checks.
These measures helped keep the human in the loop and prevented the AI agents from stopping too early.
Written by urgent.news from Dev.to's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
