Implementing DCGAN Inference on a RISC-V MCU

Using the CH32H417 RISC-V MCU, a DCGAN inference engine that generates 64x64 resolution images was implemented in pure C. It is notable for running without external libraries such as TFLite or CMSIS NN.

The key technical features are as follows:

• Model optimization: 12.6M parameters are compressed using int8 per channel quantization.
• Memory management: In the constrained 512KB SRAM environment, intermediate activations are stored in DTCM, and layer weights are streamed from the SD card using double buffering to parallelize computation and loading.
• Performance: Generating a single image takes about 26 seconds, with the main bottleneck being SD card access speed rather than computation.
• Quantum random number usage: Quantum random numbers generated via Quantum Entropy Sampling are used to generate the latent vector (z vector).