ARM NSS – neural resolution scaling technology for low end devices

ARM will start integrating NSS neural supersampling technology into its processors in 2026, which is closer in approach to AMD FSR4, as it requires an integrated AI compute unit.

The company introduced NSS as an evolution of the previously announced ARM Accuracy Super Resolution (ASR), based on AMD FidelityFX Super Resolution 2.0. The new version is accelerated by the NPU and is designed to scale resolution on low-power devices.

Starting next year, new ARM chips will begin to install NPU accelerators with a performance of more than 10 TOPS/W for AI workloads. NSS replaces a set of fixed supersampling algorithms with a lightweight AI model capable of scaling a picture from 540p to 1080p in about 4 ms per frame. This reduces the load on the GPU by up to 50% compared to native rendering.

The technology inherits the capabilities of ASR, already used in Fortnite and Infinity Nikki, and improves image clarity, stability, as well as the processing of shadows, small geometry and particle effects. NSS runs on a shared memory GPU, uses INT8 quantization, data accumulation through training, post-processing and pixel filtering without reactive masks.

The Neural Graphics Development Kit is available for developers, which includes a plugin for Unreal Engine, Vulkan emulation, profiling tools, open models on GitHub and Hugging Face, as well as an Arm machine learning extension for Arm. Project partners include Epic Games, Tencent, NetEase, Sumo Digital and Enduring Games.