Henry Ko

henryhmko.github.io

I enjoyed reading DeepSeek's NSA and I thought it would be an interesting challenge to implement and optimize it for TPUs. I put together a colab notebook containing the vectorized JAX baseline of NSA, Pallas kernels, and profiling code.

Popular posts: 1
Total points: 451
HN comments: 86
Rank: #5,810 personal

Visit henryhmko.github.io → On Hacker News Jun 2025 – Jun 2025

Hardware & Electronics 1 AI & Machine Learning 1

Most points Most discussed Most cited Newest

TPU Deep Dive

2025451 points86 comments AI & Machine Learning details

I've been working with TPUs a lot recently and it's fun to see how they had such different design philosophies compared to GPUs.