Boost ML Research with TPU Research Cloud

Abstract

How to get free access to Cloud TPUs through Google's TPU Research Cloud. It covers tensors and matrix multiplication, TPU architecture from systolic arrays to racks, training on MNIST across CPU, GPU, and TPU, where to get TPUs for free, how to apply to the program, and what that access would otherwise cost.

Outline

1. Tensors and matrix multiplication
2. TPU architecture: chips, TensorCores, MXUs, and systolic arrays
3. How TPUs connect: racks, ICI, and OCS
4. TPU v4 against the NVIDIA A100
5. Training on MNIST: CPU, GPU, and TPU
6. Where to get TPUs: Colab and Kaggle
7. TPU Research Cloud: how to apply

Key takeaways

• TPU v4 beats the NVIDIA A100 on training speed and power use
• TPUs work with the major deep learning frameworks
• Applying to the TPU Research Cloud can save thousands of dollars in compute