Making Deep Learning Go Brrrr from First Principles

(horace.io)

25 points | by tosh 1 hour ago

3 comments

  • tosh 53 minutes ago
    > in the time that Python can perform a single FLOP, an A100 could have chewed through 9.75 million FLOPS

    wild

    • patmorgan23 8 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 27 minutes ago
      This statement makes zero sense
    • xyzsparetimexyz 44 minutes ago
      Single core vs multi core accounts for much of this
      • cdavid 20 minutes 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 34 minutes 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
  • noosphr 51 minutes 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 42 minutes 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 37 minutes 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.