Unverified50% confidenceFactExact time
The original AlphaZero paper used 5,000 TPUs for self-play and 64 TPUs for training.
1
Sources
50%
Confidence
Long-term
Relevance
8/6/2026
First Seen
Sources
nanoAlphaZero：单文件JAX实现24小时训练2700分棋力
Related Claims
UnverifiedGoogle 的 TPU 引入了 bfloat16 格式（指数8位、尾数7位），牺牲精度换取更大动态范围，使训练无需损失缩放即可稳定63% similarUnverifiedThe AlphaZero reproduction used JAX and Flax to build a reinforcement learning training pipeline, conducting self-play training from scratch via multiple TPU Pods.63% similarUnverifiedThe original DETR requires 500 epochs of training on the COCO dataset (approximately 118,000 training images) to achieve satisfactory performance.62% similarUnverifiedComplete HRT comprises five core components: awareness training, competing response training, motivation enhancement, generalization training, and social support, with multiple randomized controlled trials confirming remission rates of up to 60%–70% for trichotillomania.61% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/699039API
curl https://kongchang.com/api/v1/knowledge/claims/699039MCP
get_claim(id=699039)