Chip Design 40-Second Closed Loop: LLM Generation Meets Instant Verification

Chip Design 40-Second Closed Loop: LLM Generation Meets Instant Verification

Two old paths, one deadlock

Using AI to write Verilog now basically comes down to two paths:

1. Pure LLM generation: the code looks right but doesn't compile—the registers produced by hallucination are not recognized by synthesis tools;

2. Traditional EDA (Vivado/DC and the like): closed-source, expensive, no AI assistance—RTL is still written by humans.

ChipWise takes a third path: the LLM handles generation and understanding, while the open-source EDA toolchain handles verification—generation means verification; anything that fails is automatically repaired until compilation, Lint, and simulation all pass.

8-step closed loop, ~40 seconds end-to-end

步骤能力执行引擎实测延迟
① 工具链检测EDA 健康检查本地超时保护
② 规格识别identifyOllama 端侧~2s
③ RTL 生成create云端强模型~20s
④ 编译+Lintiverilog + Verible本地~8s
⑤ 规则检查47 条(含 12 条 IEEE)本地~1s
⑥ 关系分析模块依赖+信号链本地关系引擎~1s
⑦ 仿真验证testbench 执行iverilog+VVP~8s
⑧ 时间线汇总全流程追溯本地~5s

From a natural-language specification to RTL that passes compilation, Lint, and simulation—in 40 seconds. The toolchain (Icarus Verilog / Yosys / Verible / riscv-gcc) runs entirely locally.

Measured hard data (RV32I five-stage pipeline POC)

  • iverilog compilation: 0 errors
  • style="padding-left:20px;margin:12px 0;">
  • verible lint: 0 errors
  • style="padding-left:20px;margin:12px 0;">
  • yosys synthesis succeeded: 4,069 cells / 992 flip-flops

Limitations, honestly.

Step ③ RTL generation uses a strong cloud model—the on-device 7B model proved insufficient in testing, and we won't force it. In the entire closed loop, the on-device side handles recognition, rule checking, relationship analysis, and simulation verification (spanning 7 steps), while strong generation is left to wherever it can be properly delivered. This is not a "fully on-device" promotional video; it is a workable engineering trade-off.

BossAgents (Left and Right Hands): ChipWise Engine, with all six capabilities (recognition/creation/repair/comparison/optimization/generation) fully implemented, a work validated at the HICOOL semifinals.


All data in this article can be traced in the BossAgents technical documentation library · Left and Right Hands · Boss's private AI agent

← 返回案例列表

左帮右臂工业智能体,把刚读到的方法跑起来

微信扫小程序、打开网页版 ylxt.chat,或下载桌面客户端,三种方式任选,马上用上。

🌐 网页版 ylxt.chat 💻 桌面版下载 📞 电话咨询
左帮右臂小程序码
微信扫码,体验小程序
分享:
🤖 Try Now →
🤖
🎁