GeoSeqBuilder

Training compute

6.5×10¹⁸ FLOP

Published

Sep 19, 2024

GeoSeqBuilder is an AI model developed by Peking University (China), first published in September 2024. It works in the biology domain, on tasks such as protein design and protein generation.

Training it took an estimated 6.5×10¹⁸ FLOP of compute (estimation method: hardware). It was trained on roughly 35.3M datapoints. Training ran on 1 NVIDIA A30 PCIe for about 30 hours.

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

Full record

Organization: Peking University
Country of organization: China
Domain: Biology
Task: Protein design, Protein generation
Training compute: 6.5×10¹⁸ FLOP
Compute estimation method: Hardware
Dataset size: 35.3M
Training hardware: NVIDIA A30 PCIe
Chips used: 1
Training time: 30 h
Training power draw: 179 W
Model accessibility: Open weights (non-commercial)
Open weights: Yes
Citations: 1
Epoch confidence: Confident