How PhoenixGo Turned AlphaGo Zero into a Champion AI Using Cloud Resources

After DeepMind published the AlphaGo Zero paper, the WeChat resource‑scheduling team built a large‑scale cloud platform that used idle backend compute resources for machine‑learning inference and training. Several engineers joined the PhoenixGo project to turn the paper’s algorithm into a living AI program with professional‑level Go strength.

Starting at the end of January, PhoenixGo, under the nickname “BensonDarr” ("Golden Retriever"), played on Tencent’s Wildfox Go platform against top professional players, AI opponents, and Go enthusiasts, achieving more than 200 consecutive wins by the end of April and winning the 2018 World AI Go Championship in Fuzhou.

AlphaGo Zero requires massive compute power to generate game records. PhoenixGo reused thousands of idle CPU servers from the WeChat backend, a scale of resources unavailable to most users. To enable other researchers to explore Go AI, to help players study AI moves, and to let hobbyists enjoy professional‑level play on ordinary computers, the team decided to open‑source the PhoenixGo match engine and trained model.

After a week of code cleanup, the open‑source release includes the PhoenixGo match engine source code and a 20‑block model. The training code remains closed because it is tightly coupled with the internal cloud platform. The released code and model can provide strong professional‑level strength on a single GPU block and achieve super‑human performance on multi‑GPU or multi‑machine setups. Their Wildfox Go account “Golden Retriever test” runs on a Tesla P40 with performance comparable to a GTX 1080 Ti, yielding a very high win rate against human players.

The team thanks DeepMind for publishing the AlphaGo Zero paper, which allowed a non‑Go‑expert technical team to develop PhoenixGo, and acknowledges the contributions of the community members and AI programs that helped the project succeed.