Developer adds GPU backends to pure-C TTS engine, benchmarks on rented Mac

A developer added opt-in Metal and CUDA backends to a pure-C text-to-speech engine, achieving real-time factors as low as 0.36 on M2 Pro and 0.44 on RTX 4060-class GPUs. The key architectural insight was keeping all model weights and operations resident on the GPU to avoid PCIe round-trip overhead. The developer also rented a Mac mini M2 Pro by the hour to benchmark cross-platform performance, using runtime shader compilation to avoid rebuilds.