The GPU wins the microbenchmark and loses in production.
| VAD on CPU | VAD on GPU (Quadro P1000) | Winner | |
|---|---|---|---|
| Isolated, per 1 s of audio | 1.80 ms | 1.38 ms | GPU, 24% faster |
| Production, avg per call | 13.1 ms | 18.7 ms | CPU, 30% faster |
| Production, worst call | 35 ms | 57–107 ms | CPU |
| BirdNET-Go CPU use | 25.1% of a core | 25.1% of a core | tie |
This document presents technical findings from a ground-up port of GPU-accelerated shortest-path graph algorithms (Bellman-Ford, Delta-Stepping, SPFA) from NVIDIA CUDA C++ to Apple Metal Shading Language (MSL) on the M4 architecture. The target application is PCB autorouting (OrthoRoute), where graphs contain 2,000 to 401,800 nodes and up to 2 million edges. The final Metal implementation achieves 3.7x higher throughput than an RTX 3060 on the largest tested graph, with 111.4 billion edges per second and 831.5 GB/s effective memory bandwidth.
Five critical discoveries emerged during the port. Each required solving a problem that has no direct CUDA equivalent and limited public documentation.
Run LLMs at 0% CPU on the GPU and object detection in 206ms on the NPU — both on the same Orange Pi 4A. This guide covers building Mesa + MNN from source, fixing the broken NPU driver, and running both accelerators in parallel.
| Component | Spec | Status |
|---|---|---|
| Board | Orange Pi 4A | |
| SoC | Allwinner T527 (8× Cortex-A55 @ 1.8GHz) |
The Epson TM-M30II has a 4KB receive buffer. Send data too fast → buffer overflow → gibberish and crash. Send data too slow → visible horizontal banding where the print head paused.
Think: pouring water into a cup with a hole in the bottom. Too fast overflows. Too slow leaves pour-marks.
The fix: make each raster strip fit inside the 4KB buffer.
Goal: observe a real 30-pin auth sequence end-to-end, then be able to move the auth IC to your own board.
Teuta DT-1 is an open-hardware iPod digital transport. Goal: extract Red Book digital audio (16-bit/44.1kHz PCM) from an iPod Classic via its 30-pin male dock port, route it through a hi-fi DAC (DACBerry ONE+ — WM8804 + PCM5142), output analog.
Purpose: Interface between Apple 30-pin accessories (MB129LL/A cables, docks) and Raspberry Pi for MFi chip authentication analysis, digital audio extraction, and accessory feature testing.
Design Approach: Two-board system