Imagination 白皮书解析 GPU 原生 Neural Super Resolution 神经超分技术

Insight into how neural graphics can improve image quality and performance while simplifying deployment on power- and bandwidth-constrained devices.

As AI becomes an increasingly important part of the graphics pipeline, Neural Super Resolution (NSR) offers a new approach to generating higher-quality images without the cost of rendering every pixel natively. This white paper explores how Imagination’s NSR technology uses neural reconstruction to upscale lower-resolution frames into higher-resolution outputs, combining current-frame data, temporal history and learned inference to deliver high visual quality with reduced bandwidth, power and memory demands. Unlike approaches that rely on separate AI accelerators, NSR runs directly within the GPU, enabling graphics, compute and AI workloads to work together in a unified execution environment.

Readers will learn the principles behind neural super resolution, the role of Converged Acceleration in modern GPU design, and how Imagination’s E-Series architecture integrates matrix acceleration and neural processing into the graphics pipeline. The paper also explains the importance of data locality, microtiling and Self-Compression in achieving efficient neural rendering, while examining the performance, bandwidth and energy benefits of GPU-native super resolution. For graphics architects, developers and SoC designers, it provides insight into how neural graphics can improve image quality and performance while simplifying deployment on power- and bandwidth-constrained devices.

Read more here.

Fig.1: Projected Doom Eternal performance and bandwidth improvements using NSR. Projected estimates for Doom Eternal compare native 4K rendering with 1080p rendering reconstructed to 4K using NSR on an EXD-64-2048 MC1 GPU modelled at 1 GHz. These figures should be treated as workload- and configuration-specific projections. Source: Imagination.