About

I'm Karan, a graduate student in CS at the University of Massachusetts, Amherst.
These are notes on things I find interesting.
Reach me at karanshah1698 [at] gmail [dot] com.

Recent

• Optimizing a GEMM from First Principles

  Within 95% of Intel MKL (single-threaded).

  
  

• An Opinion on Opinions

  Some "preferences" masquerade as lack of exposure.

  
  

• Embarrassingly Parallel Reduction in CUDA

  A step-by-step guide on turning simple math into a flex.

  
  

• 'Notes' of a Launch

  Hunting down music from SpaceX's Starship Flight Test recaps.

  
  

• The Blessing of Dimensionality

  'nuff said about the curse of dimensionality

  
  

• Graph Search

  Time for an algorithm shopping spree.

  
  

• PyTorch: I’m Fast, JAX: You Call That Fast?

  A recipe to train Object Detection Transformers (really) fast.

  
  

• Data Parallelism using Standard Ethernet

  Bag-of-tricks for multi-node training for the GPU Poor.

  
  

• The Algorithm

  Elon's 5-step algorithm for building great technology.

  
  

• Aster DM, Intel, CARPL collaborate for 'Secure Federated Learning Platform'

  Redirects to an external URL.