GPT-NeoX-20B

Training compute

9.3×10²² FLOP

Parameters

20B

Published

Feb 9, 2022

GPT-NeoX-20B is an AI model developed by EleutherAI (United States), first published in February 2022. It works in the language domain, on tasks such as language modeling/generation.

Training it took an estimated 9.3×10²² FLOP of compute (estimation method: hardware). The model has 20,000,000,000 parameters. It was trained on roughly 341.2B datapoints. Training ran on 96 NVIDIA A100 SXM4 40 GB for about 2.2K hours. The compute alone is estimated at $184K in 2023 dollars.

Access: Open weights (unrestricted). Its weights are openly available. The reference paper has 998 citations. Epoch AI rates the confidence of this record as confident.

Full record

Organization: EleutherAI
Country of organization: United States
Domain: Language
Task: Language modeling/generation
Training compute: 9.3×10²² FLOP
Compute estimation method: Hardware
Parameters: 20,000,000,000
Dataset size: 341.2B
Training hardware: NVIDIA A100 SXM4 40 GB
Chips used: 96
Training time: 2,160 h
Chip-hours: 207.4K
Training power draw: 77.3 kW
Training cost (2023 USD): $184K
Numerical format: FP16
Model accessibility: Open weights (unrestricted)
Open weights: Yes
Citations: 998
Epoch confidence: Confident

Benchmark results

01	PIQA	Score	0.77
02	HellaSwag	Overall accuracy	0.7
03	WinoGrande	Accuracy	0.66
04	BoolQ	Score	0.65
05	ARC AI2	Challenge score	0.41
06	OpenBookQA	Accuracy	0.39