VerilogAI

🚀 Inspiration

Writing Verilog (and other HDLs) often involves a ton of repetitive boilerplate—module headers, signal declarations, resets, clock generators, testbench scaffolding—and there isn’t yet a dedicated open-source “copilot” for hardware description languages. We set out to bring the same productivity boost that AI coding assistants have given to software engineers into the world of digital design.

✨ What It Does

1. Module Editing
   Write your Verilog modules in a full-featured Monaco-based editor.

2. AI-Generated Testbenches
   Hit Generate Testbench and get a complete, runnable testbench (with $dumpfile / $dumpvars) courtesy of a Vertex AI model.

3. Headless Simulation
   Run Icarus Verilog (iverilog + vvp) under the hood—no command line needed—and capture VCD waveform dumps automatically.

4. In-Browser Waveform Viewing
   View your VCD output right in the browser via embedded GTKWave (noVNC/Xvfb) or, in the future, a pure-JS WaveDrom pane.

5. Inline AI Suggestions
   Ghost-text completions pop up as you type common HDL patterns—Tab to accept, Esc to dismiss.

🏗 How We Built It

• Frontend

  • React + TypeScript + Chakra UI
    
  • Monaco Editor for code editing
    
  • Axios for talking to our backend APIs

• AI Layer

  • FastAPI backend
    
  • Vertex-AI model (Codestral-2501) for both inline completions & testbench generation hosted on Google Cloud Platform

• Simulation

  • Icarus Verilog inside Docker
    
  • Shared VCD volume for cross-container waveform access

• Waveform Viewer

  • GTKWave via noVNC + Xvfb/Fluxbox
    
  • Future: migrate to a JS-native WaveDrom component

• Version Control

  • GitLab for version control and hosting the project publicly

• Deployment

  • Docker Compose orchestrates:
    
  • PostgreSQL (for other app services)
    
  • FastAPI backend
    
  • React frontend
    
  • GTKWave container
    
  • Traefik handles routing & TLS

🛠 Challenges We Ran Into

• CORS & Networking
  Cross-container calls to fetch VCD files required careful CORS and volume sharing.

• Filesystem Boundaries
  Moving VCD dumps between tmpfs and host volumes forced use of cross-device-safe methods.

• Headless GTKWave
  Running GTKWave under Xvfb + noVNC, then embedding it securely in our dashboard.

• Model Constraints
  Had to rely on an off-the-shelf model (Codestral-2501) rather than a custom-trained Verilog-specific model on GCP.

• Security
  Current simulation tab still streams the whole Linux desktop. We’ll lock this down to only show the GTKWave pane.

🏅 Accomplishments

• One-Click Flow
  Module edit → AI testbench → simulate → waveform, all in under 10 seconds.

• First-Class HDL Copilot
  Seamless inline AI suggestions in Monaco for Verilog—a novel integration in the open-source world.

• Fully Containerized Demo
  Open-source, Docker-Compose setup anyone can spin up locally or in the cloud.

📚 What We Learned

• The critical role of CORS and network policy when splitting services.
• Crafting effective prompt engineering to yield valid Verilog.
• Embedding legacy GUI apps (GTKWave) in a modern web UI.

🔮 What’s Next for VerilogAI

1. Embedded GTKWave Pane
   Host only the GTKWave window (via GTKPlug or focused X11) as a dockable panel—no full desktop stream.

2. Pure-JS Waveform Renderer
   Swap out noVNC for a browser-native viewer (WaveDrom or React component) for faster, tighter integration.

3. Custom Verilog AI Model
   Train and deploy our own domain-specific model to deliver even more accurate HDL completions and testbenches.

Built With

• docker
• fastapi
• gcp
• gitlab
• javascript
• python
• react
• typescript