Making Deep Learning Go Brrrr From First Principles
Why I kept it
I like it as a practical explanation of memory-bound versus compute-bound situations and what each implies.
François Paupier
horace.io2022Public source
Read the article
In this reference
Horace He splits deep learning performance into compute, memory bandwidth, and overhead, and shows which changes help in each regime.
From the article
If you want to speed up your deep learning system, the most important thing is to understand what the bottleneck in your model is.
Horace He
Cite the original work
Horace He (2022). Making Deep Learning Go Brrrr From First Principles. horace.io. https://horace.io/brrr_intro.html
View BibTeX
@misc{making_deep_learning_go_brrrr_from_first_principles,
author = {Horace He},
title = {{Making Deep Learning Go Brrrr From First Principles}},
howpublished = {horace.io},
url = {https://horace.io/brrr_intro.html},
year = {2022}
}
