Do you have a few minutes for an ImageNet training?
Graphcore introduced IPU (Intelligence Processing Unit) a few years ago, last year the company competed successfully on MLPerf (the competition of AI chipmakers). In the image classification track, where the ResNet50 model is trained on the ImageNet dataset, the training finished in 4 minutes.
This presentation will give some information about the IPU and how is it different from the other alternative devices. What does it take to speed up the training significantly? I will showcase the most surprising modifications in the training process and the challenges with scaling it.
Artificial Intelligence Engineer, Graphcore
Bence is working on Computer Vision applications, which are optimised for Graphore’s IPU accelerators. As one of the first users of IPU specific PopTorch (PyTorch for IPUs) has first-hand experience in developing and testing a new Deep Learning framework. Prior to that collected experience in autonomous driving and NLP solutions at various companies. Bence has backgrounds in Computer Engineering and Machine Learning from BME, AIT and the Technical University of Munich.