Fujitsu: Deep Learning-ResNet-50 Learning Process Complete in 75 Seconds
Fujitsu Laboratories: Highly Efficient Distributed Parallel Processing
Fujitsu developed a technology to accelerate deep learning software and introduced it to the National Institute of Advanced Industrial Science and Technology.
Using the AI Bridging Cloud Infrastructure, we measured the time required for learning, and achieved the world’s fastest speed.
In recent years, with the spread of deep learning, an algorithm that can perform learning processing at high speed is required.
In the past two years, the speed has been increased by 30 times the original speed.
Measuring Deep Learning Processing Speed:
In ResNet-50 (Benchmark Image Recognition Neural Network), the competition time to compete for the image recognition accuracy ‘Learning time in ILSVRC 2012 (Note 5)’ is compared.
Based on HPC development technology, Fujitsu Laboratories has developed technology to extend the amount of computation per GPU, and has realized highly efficient distributed parallel processing.
AI Bridging Cloud Infrastructure Measurement: AI Bridged Cloud
In benchmarking execution, we use GPU for ABCI · 2,048.
As a result of measurement with deep learning software, it is 30 seconds or more faster than the conventional fastest recording.
It is confirmed that learning is completed in 74.7 seconds, which is the world’s fastest (Note 6).
Fujitsu