Lecture 16
Domain-specific architectures. Tensor Processing Unit.
Lecture
Outline
- Performance limitations of modern hardware
- Domain-specific architectures
- Neural networks
- Tensor processing unit
- Systolic arrays
Workshop
- Cloud TPU beginner’s guide
- How to take advantage of GPUs and TPUs for your ML project (Coding TensorFlow)
- TPU Lab in Google Colab
Homework
Play with TPU at Google Colab.
References
- Domain-specific architectures. Chapter 7 in [CAQA].
- TPU (Wikipedia)
- TPU (Google)
- Cloud TPU beginner’s guide
- Dave Patterson. Evaluation of the Tensor Processing Unit. Video. Slides.
- Dave Patterson et al. In-Datacenter Performance Analysis of a Tensor Processing Unit.
- Onur Mutlu. Computer Architecture - Lecture 27: Systolic Arrays. Video
- A Domain-Specific Architecture for Deep Neural Networks
- Google supercharges machine learning tasks with TPU custom chip
- Systolic array (Wikipedia)
- TensorFlow (Wikipedia)
- Convolutional neural network (Wikipedia)
- Multilayer perceptron (Wikipedia)
- Recurrent neural network (Wikipedia)
- Application-specific integrated circuit (ASIC) (Wikipedia)