GLM-130B

Training compute

3.5×10²³ FLOP

Parameters

130B

Published

Aug 4, 2022

GLM-130B is an AI model developed by Tsinghua University (China), first published in August 2022. It works in the language domain, on tasks such as language modeling/generation and translation.

Training it took an estimated 3.5×10²³ FLOP of compute (estimation method: operation counting,hardware). The model has 130,000,000,000 parameters. It was trained on roughly 152B datapoints. Training ran on 768 NVIDIA A100 SXM4 40 GB for about 1.4K hours. The compute alone is estimated at $820K in 2023 dollars.

Access: Open weights (non-commercial). Its weights are openly available. The reference paper has 1,264 citations. Epoch AI rates the confidence of this record as confident.

Full record

Organization: Tsinghua University
Country of organization: China
Domain: Language
Task: Language modeling/generation, Translation
Training compute: 3.5×10²³ FLOP
Compute estimation method: Operation counting, Hardware
Parameters: 130,000,000,000
Dataset size: 152B
Training hardware: NVIDIA A100 SXM4 40 GB
Chips used: 768
Training time: 1,440 h
Chip-hours: 1.1M
Training power draw: 615.7 kW
Training cost (2023 USD): $820K
Numerical format: FP16
Model accessibility: Open weights (non-commercial)
Open weights: Yes
Citations: 1,264
Epoch confidence: Confident