SiliconGym: A Versatile Reinforcement Learning Environment for Agentic EDA
Organizations: University of California San Diego, La Jolla, CA, USA · Georgia Institute of Technology, Atlanta, GA, USA · NVIDIA, Santa Clara, CA, USA
Abstract
Reinforcement learning (RL) with electronic design automation (EDA) feedback is advancing large language models for hardware design and verification. Extending these gains to hardware-engineering agents requires environments that span tasks, preserve state across tool calls, and support large hierarchies alongside parallel rollouts. We present SiliconGym, the first versatile, cross-task and cross-toolchain RL environment for hardware design and verification. By decoupling versioned workspaces, budgeted execution, and acceptance-gated scoring, SiliconGym scales from module-level RTL, testbench, and PPA tasks to multi-round repository repair and full-hierarchy SoC flows. Streaming dispatch overlaps tool checks with generation, finishing a 256-answer RL batch 21.9% faster than native ChipSeek. Using interchangeable open-source and commercial backends, we show that open-source-trained functional correctness transfers cleanly to commercial simulation, while paired netlist evaluation reveals when open-source PPA gains survive commercial synthesis.
Figures & tables
| Category | Task / workload | Scale / interaction | Evaluation stages | Backends | Catalog / execution scope |
| Training | RTL generation / CodeV | Module / parallel RL | Simulation | OSS, Comm. | 3,133 defs.; updates |
| PPA optimization / ChipSeek | Module / parallel RL | Sim., synthesis | OSS, Comm. | updates; 32 slots | |
| Cases | Repository repair / HWE | Repository / Bash agent | Build, regression | OSS | 5 repairs; 1 Record each |
| Feedback RTL repair | Module / multi-round agent | Sim., waveforms | OSS | Arbiter + 3 WaveDebug repairs | |
| Formal RTL / VeriThoughts | Module / controls, repair | Bounded equivalence | OSS | 5 qualified; 1 two-round repair | |
| Testbench / CorrectBench | Module / TB controls | Sim. (golden/mutant) | OSS | 5 qualified cases |
| Repository | Lang. / domain | Edited files | Bash cmds | Targeted regression |
|---|---|---|---|---|
| Ibex | SV / RISC-V | 8 | 17 | Debug-mode PMP access |
| CVA6 | SV / RISC-V | 1 | 8 | Compressed decode legality |
| Caliptra | SV / RoT | 1 | 7 | Key-vault error retention |
| Rocket Chip | Chisel / RISC-V | 1 | 16 | JTAG asynchronous reset |
| XiangShan | Chisel / RISC-V | 1 | 13 | Virtual interrupt selection |
| Scoring backend | |||
|---|---|---|---|
| Model | OSS (%) | Comm. (%) | (p.p.) |
| Initial | 106.04 | 101.56 | +4.48 |
| A (OSS-trained) | 106.11 | 101.19 | +4.93 |
| B (comm.-trained) | 104.34 | 100.10 | +4.24 |
| (p.p.) | +1.78 | +1.09 | +0.69 |
Electronic Design Automation
Electronic Design Automation uses algorithms and software to design, verify, optimize, and manufacture electronic circuits and systems.
See all 256 papers