Down with NVIDIA! Running Qwen on a $50 FPGA: Building the 'KEDR' TPU

The author shares an ambitious experiment in developing a custom AI processor at home. Instead of expensive NVIDIA solutions, the enthusiast utilized a budget-friendly $50 FPGA board with 4GB of memory. By connecting it to a standard mini-PC, they successfully ran the Qwen language model. The project, dubbed 'KEDR,' is based on the openTPU architecture and demonstrates the feasibility of implementing specialized AI computing systems on accessible hardware. In the first part of the series, the author details the hardware selection process, the technical challenges of integrating the FPGA with a computing node, and the initial performance results. This experiment highlights the growing interest in custom AI accelerators and the accessibility of programmable logic technologies for independent developers seeking to bypass the limitations of traditional GPUs.