Engineer Builds GPU from Scratch in Two Weeks

Adam Majmudar, an engineer at a web3 development company, successfully built a GPU from scratch in just two weeks, sharing his journey on Twitter/X and GitHub. The project, named "tiny-gpu," has gained over 5,300 stars and demonstrates the core concepts of GPU architecture through a simplified design.

Majmudar began by learning the basics of GPU architecture, including CUDA frameworks, SIMD programming, and core components like global memory, compute cores, hierarchical caches, and memory controllers. He then designed a minimal GPU architecture focusing on parallelization, memory access, and resource management, targeting GPGPU functionality for machine learning applications.

The project involved creating a custom 11-instruction ISA inspired by LC4, writing matrix math kernels, and building the GPU using Verilog. With guidance from hacker George Hotz, Majmudar overcame challenges such as implementing asynchronous memory and memory controllers, and refining the execution scheduling. The final design was verified using OpenLane EDA and is set to be fabricated through Tiny Tapeout 7.

This achievement highlights the growing accessibility of chip design through open-source tools and agile methodologies, inspiring further exploration in DIY chip development.