game-engines

Arm's new mobile GPU renders a quarter of the frame and infers the rest

September 8, 2026by GameDevBuddy Team

Arm's Mali G2-Ultra NX puts neural accelerators inside the shader cores, and publishes the networks openly so studios can retrain them on their own art.

Arm announced the Mali G2-Ultra NX today at its Arm Everywhere China event in Shanghai, the first Mali GPU with neural accelerators built into the shader cores. For mobile developers it means a rendering budget that no longer scales with pixels: Arm's demo game renders one quarter of a frame conventionally and reconstructs the rest, then renders the following frame entirely on the neural hardware.

The GPU is the graphics half of Arm CSS for Mobile 2, the platform Arm announced alongside it. Three technologies ride on the new accelerators, all described on Arm's newsroom: Neural Super Sampling reconstructs a high-resolution image from a low-resolution render, Neural Frame Rate Upscaling generates intermediate frames, and Neural Super Sampling and Denoising pairs upscaling with ray reconstruction for ray-traced scenes. Because the accelerators sit inside the shader cores rather than beside them, neural work shares the GPU's memory system and caches instead of shuttling data across the chip.

The numbers Arm quotes come from Neural Dawn, a showcase game built with Sumo Digital: up to four times the performance efficiency and up to 70 percent less external memory traffic than native rendering. Arm's claim is that this headroom makes techniques such as Unreal Engine's MegaLights practical on a phone. Separately, the new execution engine — Arm calls it the largest Mali instruction set change in seven generations — delivers up to 24 percent higher benchmark performance and 14 percent higher performance in non-AI gaming workloads over the previous generation, and the third-generation ray tracing unit cuts DRAM traffic on ray tracing benchmarks by up to 13 percent.

The part worth reading closely is how Arm has positioned this against Nvidia. Speaking to GamesIndustry.biz, Arm GPU product manager Deyan Lazarov said the networks are convolutional rather than generative — "they don't imagine anything, they don't create anything" — and that all of them are published openly on GitHub and Hugging Face. Studios can capture frames from their own game and retrain the network on that data, which is the difference that matters if your art direction is stylised. A generic upscaler trained on photoreal content will soften a hand-painted character or a hard-edged robot; one retrained on your own frames will not, and the retraining happens on your data, in your hands.

Tooling is the usual catch. Arm has plug-ins for Unreal available now, with Unity described only as something the company is looking at, and an SDK for teams on custom engines. Lazarov put integration into an already-shipping title at around three months. No handset has been confirmed as carrying the chip; the previous Mali G1-Ultra turned up in devices from Xiaomi, Vivo and OnePlus.

#unreal-engine#mobile#rendering