GateGPT: 56k tokens per second Transformer (KV cache) on FPGA at 80 MHz

A researcher creates a custom chip that can run a full Transformer model at a high speed.