Engineer 'builds a GPU from scratch' in two weeks — process much harder than he expected

An engineer has shared his journey in “building a GPU from scratch with no prior experience.” As with his prior project of designing a CPU from scratch, Adam Majmudar took just two weeks to complete this cerebral feat. In a Twitter/X thread Majmudar takes us through the process, step-by-step, and admits GPU designing was a much harder task than expected. To be clear, the current conclusion of the project is a chip layout in Verilog which was finally passed through OpenLane EDA software to verify it. However, the GPU is going to be submitted for tapeout via Tiny Tapeout 7 so is destined to become a physical chip in the coming months.

Above you can see the flow of tasks Majmudar worked through to design his GPU. Yet, as a ‘from scratch’ project, a lot of study and thought was required even before the first step was tentatively taken. Last time we highlighted the engineer’s concerns that GPUs would be a relatively difficult field of study, due to the dominance of proprietary tech, as that prediction came true.

Through several iterations of the above architecture, Majmudar decided to focus on general-purpose parallel computing (GPGPUs) capabilities. Thus he adjusted his Instruction Set Architecture (ISA), which features just 11 instructions, to achieve this goal. Next up, the engineer wrote two matrix math kernels to run on his GPU. These matrix addition and multiplication kernels would demonstrate the key functionality of the GPU and provide evidence of its useful application in graphics and machine learning tasks.

Leave a Comment