somnial

15 karmajoined October 30, 2025On HN

Recent submissions

What happens when a GPU writes memory blog.doubleword.ai
5somnial12ddiscuss
What happens when a GPU reads memory?blog.doubleword.ai
15somnial1modiscuss
The case for disaggregated LLM serving blog.doubleword.ai
4somnial1modiscuss
On-the-fly snapshot compression for elastic inference at scale blog.doubleword.ai
6somnial1mo1 comment
NVLink, NVSwitch, and All That blog.doubleword.ai
5somnial1mo2 comments
The Anatomy of an Instruction Pipeline Hazard hiraditya.github.io
9somnial2modiscuss
Width vs. Depth: Speculating on the Margin blog.doubleword.ai
17somnial2mo1 comment
Pushing memory bound CUDA kernels past the speed of light with data compression fergusfinn.com
2somnial3modiscuss
Speculative KV coding: ~4× losslessly compressed KV cache using a small model fergusfinn.com
2somnial4modiscuss
70x faster cold(ish) starts for SGLang fergusfinn.com
1somnial4modiscuss
LLM powered data structures: A lock-free binary search tree fergusfinn.com
1somnial7modiscuss
Parallel Primitives for Multi-Agent Workflows fergusfinn.com
1somnial8modiscuss
Scheduling in LLM Inference fergusfinn.com
1somnial9modiscuss
How fast can an LLM go?fergusfinn.com
2somnial10modiscuss

All submissions and comments on Hacker News.