AlphaStar

Training compute

1.1×10²³ FLOP

Parameters

139M

Published

Oct 30, 2019

AlphaStar is an AI model developed by DeepMind (United Kingdom), first published in October 2019. It works in the games domain, on tasks such as starcraft. It counts among the frontier models: the systems trained with the most compute of their moment.

Training it took an estimated 1.1×10²³ FLOP of compute (estimation method: hardware). The model has 139,000,000 parameters. Training ran on 384 Google TPU v3 for about 1.1K hours. The compute alone is estimated at $131K in 2023 dollars.

Access: Unreleased. Its weights are not openly released. The reference paper has 4,093 citations. Epoch AI rates the confidence of this record as confident.

Full record

Organization: DeepMind
Country of organization: United Kingdom
Domain: Games
Task: StarCraft
Training compute: 1.1×10²³ FLOP
Compute estimation method: Hardware
Parameters: 139,000,000
Training hardware: Google TPU v3
Chips used: 384
Training time: 1,056 h
Chip-hours: 405.5K
Training power draw: 354.2 kW
Training cost (2023 USD): $131K
Model accessibility: Unreleased
Open weights: No
Citations: 4,093
Epoch confidence: Confident