FFN SwiGLU

Training compute

3.4×10¹⁹ FLOP

Parameters

220M

Published

Feb 14, 2020

FFN SwiGLU is an AI model developed by Google (United States), first published in February 2020. It works in the language domain, on tasks such as language modeling and question answering.

Training it took an estimated 3.4×10¹⁹ FLOP of compute (estimation method: hardware,operation counting). The model has 220,000,000 parameters. It was trained on roughly 50.6B datapoints. Training ran on 16 Google TPU v2 for about 21.85 hours.

Access: Unreleased. Its weights are not openly released. Epoch AI rates the confidence of this record as confident.

Full record

Organization: Google
Country of organization: United States
Domain: Language
Task: Language modeling, Question answering
Training compute: 3.4×10¹⁹ FLOP
Compute estimation method: Hardware, Operation counting
Parameters: 220,000,000
Dataset size: 50.6B
Training hardware: Google TPU v2
Chips used: 16
Training time: 22 h
Training power draw: 9.2 kW
Model accessibility: Unreleased
Open weights: No
Epoch confidence: Confident