Live Demo Deployment
This page describes the deployment setup behind the live demo at sungyongcho.com/gomoku (including this docs site).
Infrastructure
| Service | GCP Machine | vCPU | RAM |
|---|---|---|---|
| Minimax | e2-small | 0.5–2 | 2 GB |
| AlphaZero | c2d-standard-4 | 4 | 16 GB |
Both backends run as Docker containers on GCP Container-Optimized OS VMs. A Cloudflare Worker handles path-based routing (/minimax/*, /alphazero/*) to the respective backends.
Training used a Ray cluster with one GPU head node and CPU-only self-play workers. CPU workers generated games in parallel and sent inference requests to GPU actors through the async inference pipeline. See Training Pipeline for the cluster architecture.
For full deployment engineering details (Cloudflare Worker path routing, COS startup scripts, Artifact Registry image flow, and deployment scripts), see AlphaZero Deployment.
Response times on the live demo will vary from benchmark numbers cited in the documentation — the minimax engine in particular runs on a shared-core instance to keep hosting costs low.