The Real-Time 4K Image Super-Resolution on CPU with Hybrid of Post-Training Quantization and Quantization-Aware Training

This Real-time 4K image super-resolution (SR) is an intensive task that demands substantial computational resources, typically allocated to GPUs or ASICs. In contrast, most high-quality video transcoding workloads are executed on CPUs. This disparity necessitates the use of different hardware platforms for super-resolution and transcoding, resulting in increased complexity and elevated costs. With the advent of the new AI instructions, AMX, on the 5th generation Xeon platform, a single-socket CPU is now capable of delivering over 200 TOPs of INT8 computational power. Despite this advancement, the current quantization algorithms for SR models, including Quantization-Aware Training (QAT) and Post-Training Quantization (PTQ), fall short of expectations. In this paper, we introduce an innovative hybrid algorithm tailored for the quantization of image super-resolution models, designed to strike an optimal balance between processing speed and image quality.