Vibecoding an inference chip
Summary
Ankit Gupta details vibe-coding an FPGA to run an LLM, using AWS F2 FPGA instances and a long autoresearch loop that produced real-time throughput improvements up to 213 tokens per second. The post emphasizes a hands-on hardware-software co-design workflow, including C++/Verilog, memory architecture, prompts, and a thorough optimization log, illustrating how AI tooling and cloud FPGA resources can accelerate hardware-backed AI inference.