# Clone autoresearch and copy in the parallel experiment files
foresight of a long chain of consequences, of which very seldome any man
，推荐阅读wps获取更多信息
Portable USB AI inference accelerator. Runs selected MoE models with up to 120B total parameters, but much smaller active per-token workloads, at roughly 12–16 tok/s under short-context conditions. Longer contexts degrade sharply, with roughly 6–9 tok/s in the 8K–32K range and very high TTFT at 32K+. Requires host computer and proprietary desktop software. Uses split memory architecture across a 32GB SoC pool and 48GB dNPU pool connected over PCIe. Model support is limited to pre-optimized builds from TiinyAI’s store. Inference stack builds on PowerInfer research from SJTU IPADS.
I wanted some form of I/O co-processor in Baochip, so I studied the PIO the best way I knew how – by copying it. I forked Lawrie Griffith’s fpga_pio as a starting point, and did a whole bunch of regression testing and detail simulation to clean up all the missing corner cases. You can find what I think is fairly close to a fully spec-compliant RP2040-generation PIO core in this github repo.