README

██████╗ ██╗████████╗███████╗███████╗██╗  ██╗
██╔══██╗██║╚══██╔══╝██╔════╝██╔════╝██║  ██║
██████╔╝██║   ██║   █████╗  ███████╗███████║
██╔══██╗██║   ██║   ██╔══╝  ╚════██║██╔══██║
██║  ██║██║   ██║   ███████╗███████║██║  ██║
╚═╝  ╚═╝╚═╝   ╚═╝   ╚══════╝╚══════╝╚═╝  ╚═╝
██╗   ██╗ █████╗ ██████╗  █████╗ ██╗   ██╗
╚██╗ ██╔╝██╔══██╗██╔══██╗██╔══██╗██║   ██║
 ╚████╔╝ ███████║██║  ██║███████║██║   ██║
  ╚██╔╝  ██╔══██║██║  ██║██╔══██║╚██╗ ██╔╝
   ██║   ██║  ██║██████╔╝██║  ██║ ╚████╔╝
   ╚═╝   ╚═╝  ╚═╝╚═════╝ ╚═╝  ╚═╝  ╚═══╝

Independent research notes by Ritesh Yadav on ML performance, infrastructure, and systems: CUDA, low-latency inference, generative AI, distributed training, Kubernetes, and LLMOps.

These pages collect working notes on the GPU stack and production inference: definitions, references, and the reasoning I use when reading kernels, profilers, and serving systems. They are written for careful reading, not as product documentation.

Research Topics

This is the map of the notes. Each area links to its index page and calls out the questions it is meant to answer.

#	Research area	What it covers
01	Device Hardware	Microarchitecture, streaming multiprocessors, and the memory hierarchy
02	Device Software	CUDA programming, kernels, tiling, and device-side execution
03	Host Software	Driver APIs, runtimes, CUTLASS, and CuTe
04	Performance	Roofline analysis, memory coalescing, and bank conflicts
05	Inference Engineering	Quantization, vLLM, and low-latency serving
06	Large Language Models	Attention kernels, parallelism, and mixture-of-experts systems

Browse the Notes

The table of contents on the left is the primary index on desktop. On smaller screens, open the menu to move between areas. The previous and next arrows at the bottom of each page follow the same reading order.