Built an AI Accelerator and opensourced it. [P]

There is a huge gap in open source AI accelerators, so I implemented mine . Popular and well known ones are already legacy and doesn't support contemporary operations like Attention. Here is what makes mine special: Attention mechanism smelted directly into silicon Prototyped end-to-end on FPGA (AWS F2) Benchmarked against PyTorch -based workloads Built on the RocketChip architecture (RISC-V) Native BF16 support Up to 225× speedup on vanilla attention mechanism Up to 96× speedup on TinyBERT Up to 50× speedu