Run TinyML Models in Browser via $35 USB FPGA in Microseconds

A developer workflow called WebFPGA combines the browser's WebUSB API with Lattice iCE40-UP5K FPGA boards to run machine learning inference directly from a webpage without any additional software installation. A TensorFlow Lite model is converted to Verilog using the tinymlgen tool, then compiled into a bitstream that the browser loads onto the FPGA over USB. The entire pipeline — from a .tflite file to real-time browser inference — can be set up in under five minutes using open-source tools such as Yosys and NextPNR. Benchmarks show the FPGA approach achieves roughly 30 microseconds of average latency at just 0.5W power consumption, outperforming WebGPU, Raspberry Pi, and cloud-based alternatives in both speed and energy use. For large-scale edge deployments, the solution is estimated to cut total operational costs by over 99% compared to cloud inference services.