MnasNet-A3

Training compute

1.5×10²¹ FLOP

Parameters

5.2M

Published

May 29, 2019

MnasNet-A3 is an AI model developed by Google (United States), first published in May 2019. It works in the vision domain, on tasks such as image classification and object detection.

Training it took an estimated 1.5×10²¹ FLOP of compute (estimation method: hardware). The model has 5,200,000 parameters. It was trained on roughly 1.2M datapoints. Training ran on 256 Google TPU v3 for about 108 hours. The compute alone is estimated at $10K in 2023 dollars.

Access: Open weights (unrestricted). Its weights are openly available. The reference paper has 3,396 citations. Epoch AI rates the confidence of this record as speculative.

Full record

Organization: Google
Country of organization: United States
Domain: Vision
Task: Image classification, Object detection
Training compute: 1.5×10²¹ FLOP
Compute estimation method: Hardware
Parameters: 5,200,000
Dataset size: 1.2M
Training hardware: Google TPU v3
Chips used: 256
Training time: 108 h
Training power draw: 237.0 kW
Training cost (2023 USD): $10K
Numerical format: FP32
Model accessibility: Open weights (unrestricted)
Open weights: Yes
Citations: 3,396
Epoch confidence: Speculative