The Missing 2x: Why Int8 is Twice as Fast as BF16
From bf16 to int8 subword parallelism
From Scratch
I build software from scratch and share my learnings! I especially enjoy topics around compilers, distributed systems, GPU programming, and Linux.
Subscribe:.rss.atom.json.md.m3u.pls
Live Last read · last published · next check
Latest episodes
Paper Journal: Disaggregated Serving
Achieving consistent online LLM inference performance
Optimizing matrix multiplication
Discovering optimizations one at a time
Linux container from scratch
Let's build a minimal container step-by-step in a terminal
Primer on Linux container filesystems
Building a container filessytem by hand
Inference Engine: Accelerating with CUDA
Still crunching numbers, but faster.
Inference Engine: Optimizing Performance
How I made my inference engine 15x faster
Build Your Own Inference Engine: From Scratch to "7"
Building a C++ Inference Engine from scratch
GPU Programming
Writing code for massively parallel processors
MapReduce from Scratch
Building a distributed computing framwork, step by step.