sim2gym turns a real engineering simulator into a training and testing environment for AI models.

Engineers design hardware by experimenting: they run a simulator, read the result, and adjust. sim2gym gives an AI model the same job. Our first environment, AccelGym, asks the model to design the fastest AI chip for a real neural-network layer from BERT-tiny, ResNet-18 or MobileNetV2.

How it works

1. The model gets a design problem: one layer and a chip-size budget.
2. It chooses the chip’s grid shape and dataflow, and tests ideas on SCALE-Sim, a real open-source chip simulator from Georgia Tech.
3. It can run only 6 experiments, so it must learn from each result.
4. It submits one final design. We simulate every allowed design in advance, so each answer gets an exact, fair score. No AI judge.

What we built

• A Python environment that wraps SCALE-Sim, with exact scoring and a disk cache for repeatable results
• A live web demo that streams each episode: every model message, every simulator run, and the final design next to the best one
• A task generator that creates unlimited new design problems from a seed
• A Prime Intellect verifiers environment, so open-source RL trainers can train models to experiment better
• Support for open-weight models through Amazon Bedrock, OpenRouter and Ollama
• 147 automated tests

Any simulator that is a Python function can become a new environment. Chip design is the first.