Making Deep Learning Go Brrrr from First Principles

Making Deep Learning Go Brrrr from First Principles(horace.io)

32 points by tosh 2 hours ago | 14 comments

tosh 1 hour ago |

> in the time that Python can perform a single FLOP, an A100 could have chewed through 9.75 million FLOPS

wild

patmorgan23 51 minutes ago | |

Why are we comparing a programing language and a GPU. This is a category error. Programing languages do not do any operations. They perform no FLOPs, they are the thing the FLOPs are performing.

"The I7-4770K and preform 20k more Flops than C++" is an equally sensible statement (i.e. not)

p1esk 1 hour ago | |

This statement makes zero sense

xyzsparetimexyz 1 hour ago | |

Single core vs multi core accounts for much of this

cdavid 1 hour ago | | |

Not really. GPU many cores, at least for fp32, gives you 2 to 4 order of magnitudes compared to high speed CPU.

The rest will be from "python float" (e.g. not from numpy) to C, which gives you already 2 to 3 order of magnitude difference, and then another 2 to 3 from plan C to optimized SIMD.

See e.g. https://github.com/Avafly/optimize-gemm for how you can get 2 to 3 order of magnitude just from C.

jdw64 1 hour ago |

Right now, all I know how to do is pull models from Hugging Face, but someday I want to build my own small LLM from scratch

kflansburg 38 minutes ago | |

If you aren't already aware, Karpathy has several videos that could get you there in a few hours https://www.youtube.com/@AndrejKarpathy

jdw64 36 minutes ago | | |

very thanks!

max-amb 27 minutes ago | |

If you want a written resource I have a blog post about the mathematics behind building a feed forward from scratch, https://max-amb.github.io/blog/the_maths_behind_the_mlp/. Kinda focuses on translation from individual components to matrix operations.

glouwbug 42 minutes ago | |

It’s just linear algebra. Work your way from feed forward to CNN to RNN to LSTM to attention then maybe a small inference engine. Kaparthy’s llama2.c is only ~300 lines on the latter and it pragma simds so you don’t need fancy GPUs

noosphr 1 hour ago |

>For example, getting good performance on a dataset with deep learning also involves a lot of guesswork. But, if your training loss is way lower than your test loss, you're in the "overfitting" regime, and you're wasting your time if you try to increase the capacity of your model.

https://arxiv.org/abs/1912.02292

appplication 1 hour ago | |

Generally, posting a link-only reply without further elaboration comes across as a bit rude. Are you providing support for the above point? Refuting it? You felt compelled to comment, a few words to indicate what you’re actually trying to say would go a long way.

noosphr 1 hour ago | | |

>We show that a variety of modern deep learning tasks exhibit a "double-descent" phenomenon where, as we increase model size, performance first gets worse and then gets better.