Arm launches Mali G2-Ultra NX with neural acceleration for mobile games

Arm introduced Mali G2-Ultra NX on September 8 as its first Mali GPU with dedicated neural accelerators integrated into the shader cores. The component belongs to the CSS for Mobile 2 platform and is designed to run image reconstruction, frame generation and denoising alongside conventional graphics workloads. Its commercial debut was reported in the Xiaomi 18 Fold, launched in China with the Xring O3 chip.

The announcement brings phones closer to a strategy already common in PC GPUs: render fewer pixels or fewer complete frames, then use specialized models to reconstruct part of the result. That approach can increase visual quality within a mobile device’s strict power, thermal and bandwidth limits. However, the figures released so far are Arm estimates and demonstrations, not independent tests guaranteeing the same result on every device.

What changes inside the GPU

Mali G2-Ultra NX places neural accelerators inside the shader system, sharing memory, caches and control structures with other GPU workloads. Arm has also introduced a new execution engine, described as its biggest Mali instruction-set architecture change in seven generations, and a third-generation ray-tracing unit.

Three techniques form the core of the announcement. Neural Super Sampling, or NSS, reconstructs a higher-resolution image from a lower-resolution render. Neural Frame Rate Upscaling, or NFRU, creates intermediate frames using rendered images, motion and depth. Neural Super Sampling and Denoising, or NSSD, combines reconstruction with noise reduction in ray-traced scenes.

In the manufacturer’s published tests, the new architecture delivers up to 24% higher benchmark performance and 14% higher non-neural gaming performance, both compared with the previous Mali generation. In Neural Dawn, a demonstration produced with Sumo Digital, the combined techniques reached up to four times higher performance efficiency and up to 70% lower external-memory traffic than native rendering. Those figures describe Arm-selected scenarios and should not be treated as automatic multipliers for every game.

The features will not all arrive at once

Supersampling is the first step. Chris Bergey, Arm’s executive vice president of Edge AI, told The Verge that neural frame generation is planned for early 2027 and denoising will follow later. Where Winds Meet, Infinity Nikki and Arena Breakout: Infinite are among the first games associated with the technology.

Adoption also requires work from studios. Games must provide data such as motion vectors and depth while integrating the features into their rendering engines. Arm offers a development kit, Vulkan extensions and support for Unreal Engine workflows; Tencent Games and Unity China are among the partners named by the company. Having the GPU in a phone therefore does not automatically enable the features across the Android catalog.

More displayed frames do not mean proportionally faster response

Frame generation can make animation look smoother, but some of the displayed rate comes from intermediate images rather than new, fully simulated game frames. As our guide to frame generation, FPS and latency explains, the visual increase does not reduce response time by the same proportion and can add processing to the pipeline.

That limitation does not make the approach irrelevant. On phones, reducing rendering work can preserve quality for longer before heat and power consumption force a chip to lower its clocks. The more important consequence is not simply a promise of “PC graphics,” but a hardware and software foundation that lets developers exchange part of the brute-force cost for neural reconstruction.

Arm says the technology will reach other phones, tablets and larger-screen devices in 2027, but it has not yet provided a complete global list. Until more products and games are tested, the architecture, tools and early integrations are confirmed; battery life, image quality and sustained performance will still depend on each implementation.