End of Moore’s Law

Perhaps the best source to describe what is happening with chips is Dr. David Patterson a professor at UC Berkeley and an architect for the TPU (Tensorflow Processing Unit).

Screen Shot 2020-02-18 at 11 49 01 AM

The high-level version is that CPU clock speed is effectively dead. This opens up an opportunity for new solutions. These solutions involve both cloud computing and also specialized chips called ASICs.



TPU in Production TPU Production

CPU How CPUs work

GPU GPUs work

TPU TPUs work

Sources: https://storage.googleapis.com/nexttpu/index.html

Using GPUs and JIT

Screencast Do GPU Programming with CUDA, Numba and Colab!

One of the easiest ways to use a Just in Time compiler (JIT) or a GPU is to use a library like numba and a hosted runtime like Google Colab.

There is a step by step example of how to use these operations in the following notebook (https://github.com/noahgift/cloud-data-analysis-at-scale/blob/master/GPU_Programming.ipynb)[https://github.com/noahgift/cloud-data-analysis-at-scale/blob/master/GPU_Programming.ipynb]. The main high level takeaway is that the GPU runtime in colab must be enabled.

Screen Shot 2020-02-18 at 12 44 56 PM

Next up install numba and double check the CUDA .so libraries are available.

!pip install numba
!find / -iname 'libdevice'
!find / -iname 'libnvvm.so'

You should see something like this.


Next up try one of the methods for speeding up your code.

GPU Workflow

[TO DO: Create GPU vectorize workflow diagram]


  • Topic: Go through colab example here
  • Estimated time: 20-30 minutes
  • People: Individual or Final Project Team
  • Slack Channel: #noisy-exercise-chatter
  • Directions:
    • Part A: Get code working in colab
    • Part B: Make your own GPU or JIT code to speed up a project you are working on. Share in slack and/or create a technical blog post about it.