Top all-time · #9516
Horace He
Personal blog · thonking.aiaiprogrammingsystems
Horace He writes Thonk From First Principles, a personal blog at thonking.ai about ai, programming and systems. Since 2024, 8 different Hacker News readers have linked it 10 times, most often “Matrix Multiplications on GPUs Run Faster When Given Predictable Data”. It is still writing: the last post was on 2025-03-04, and you can follow it by feed.
Too few sampled comments (2) to say how often HN recommends it.?
Why HN links it
HN readers link Horace He's posts on AI hardware and inference performance, including whether NPUs beat CPUs for AI workloads, Python versus lower-level code, and limits on LLM speed.
Summarised by Claude from a sample of 2 HN comments that link here. Each comment is marked recommending, critical or neutral (citing without judging); how this works.
ML Systems from first principles. Aims to be better than a ChatGPT summary.
Most-linked posts?
- Matrix Multiplications on GPUs Run Faster When Given Predictable Data5 people · 5 mentions · first linked 2024-04-30
- What Shapes Do Matrix Multiplications Like?3 people · 4 mentions · first linked 2024-04-01
What Hacker News says?
> but to be able to run small models with very little power usage yes But first, I should also say you probably don't want to be programming these things with python. I doubt you'll get good performance there, especially as the newness means optimizations haven't been ported well (even using things like TensorRT is not going to be as fast as writing it from scratch, and Nvidia is throwing a lot of man power at that…
Comment linking /p/what-shapes-do-matrix-multiplications
on “AI PCs Aren't Good at AI: The CPU Beats the NPU”Submitted What Shapes Do Matrix Multiplications Like? as a story · 164 points
HN discussionNot enormous without significant changes to the ML. There are two pieces to this: improving efficiency and improving flops. Improving flops is the most obvious way to improve speed, but I think we're pretty close to physical limits for a given process node and datatype precision. It's hard to give proof positive of this, but there are a few lines of evidence. One is that the fundamental operation of LLMs, matrix mul…
Comment linking /p/strangely-matrix-multiplications
on “Llama 3.1 405B now runs at 969 tokens/s on Cerebras Inference”Submitted Matrix Multiplications on GPUs Run Faster When Given Predictable Data as a story · 6 points
HN discussionSubmitted Matrix Multiplications on GPUs Run Faster When Given Predictable Data as a story · 4 points
HN discussionSubmitted What shapes do matrix multiplications like? as a story · 4 points
HN discussionSubmitted What shapes do matrix multiplications like? as a story · 2 points
HN discussionSubmitted Matrix Multiplications on GPUs Run Faster When Given "Predictable" Data as a story · 1 points
HN discussionSubmitted Why PyTorch is an amazing place to work and Why I'm Joining Thinking Machines as a story · 1 points
HN discussionSubmitted Matrix Multiplications Run Faster When Given "Predictable" Data as a story · 1 points
HN discussion