JFT

JFT is an AI model developed by Google Research and Carnegie Mellon University (CMU) (United States), first published in July 2017. It works in the vision domain, on tasks such as image classification, object detection, semantic segmentation and pose estimation. It counts among the frontier models: the systems trained with the most compute of their moment.

Training it took an estimated 8.4×10²⁰ FLOP of compute (estimation method: hardware). The model has 44,654,504 parameters. It was trained on roughly 5.5T datapoints. Training ran on 50 NVIDIA Tesla K80 for about 1.4K hours. The compute alone is estimated at $18K in 2023 dollars.

The reference paper has 2,712 citations. Epoch AI rates the confidence of this record as confident.