TinyGAN: Generative Image Synthesis on a RISC-V Microcontroller with Quantum Entropy Sampling

We present TinyGAN, a demonstrated deployment of generative adversarial network image synthesis on a resource constrained RISC-V microcontroller. A DCGAN generator with 12.6 million parameters produces 64x64 RGB images of cat faces on a CH32H417 dual core RISC-V system on chip operating at 150/300 MHz with 512 KB of shared SRAM and no external memory. Through a combination of dual core compute offloading, int8 per channel quantization, tiled weight layout optimization, double buffered SD card streaming, and cross core DTCM intermediate storage, we reduce inference time from 242 seconds to 26 seconds, a 9.3x improvement. Latent vectors are seeded from quantum vacuum fluctuation measurements provided by the Australian National University quantum random number generator. The system includes combinatorial phrase synthesis and neural text to speech audio output, creating an autonomous generative art installation requiring no network connectivity at runtime