Arm's Mali G2-Ultra NX brings AI-native neural graphics to mobile GPUs

Arm has announced the Mali G2-Ultra NX, a mobile GPU the company describes as the first AI-native Mali design. The chip places neural accelerators directly inside the shader cores, so AI-based graphics workloads run alongside conventional rendering and compute rather than on a separate block. Arm's pitch is straightforward: mobile users keep demanding richer visuals and smoother frame rates, but phone silicon cannot simply scale up the way a desktop GPU can, so the extra fidelity has to come from AI doing part of the work instead.

The announcement, published on Arm's newsroom and surfaced on Hacker News' front page, positions neural graphics — using machine learning to reconstruct detail, generate intermediate frames and refine images — as the path to desktop-class visuals inside a mobile power envelope.

Neural graphics wired into the pipeline

According to Arm, the neural accelerators share the GPU's memory system, coherent caches and control structures, which reduces data movement and improves efficiency across the pipeline. Three neural techniques ship with the design:

• Neural Super Sampling (NSS) reconstructs a higher-resolution image from a lower-resolution render, combining temporal data with integrated anti-aliasing so the GPU does less conventional rendering for the same final quality.
• Neural Frame Rate Upscaling (NFRU) generates intermediate frames from motion, depth and rendered-frame data to raise frame rates and smooth gameplay.
• Neural Super Sampling and Denoising (NSSD) pairs neural upscaling with denoising for demanding ray-traced scenes with complex lighting and shadows.

To demonstrate the pipeline in practice, Arm built a showcase title, Neural Dawn, with UK studio Sumo Digital. Arm reports that combining NFRU and NSSD delivered up to 4x higher performance efficiency and up to 70 percent lower external memory traffic compared with native rendering — gains the company says are large enough to make techniques such as Unreal Engine MegaLights practical on mobile devices.

A new execution engine and third-generation ray tracing

Alongside the neural hardware, Arm has reworked the execution engine, which it calls the biggest Mali instruction set architecture upgrade in seven generations. It provides up to twice the registers per warp, giving shaders headroom for more demanding scenes while maintaining frame rates. According to Arm, the architectural changes alone deliver up to 24 percent higher benchmark performance and 14 percent higher non-AI gaming performance compared with the previous generation, and combining the new hardware with NFRU supports gaming sessions at up to 120 FPS.

Ray tracing reaches its third generation in this design, supporting more complex lighting, shadows, reflections and geometry. Arm claims up to 13 percent lower DRAM traffic on its leading ray tracing benchmarks. The GPU also adds hardware support for Opacity Micromaps, a technique for handling intricate transparent geometry such as foliage and layered fabrics; in one Arm demo, enabling it raised frame rates by 30 percent and cut the ray tracing workload by up to 70 percent.

Tooling so developers can actually ship it

Arm says it began preparing the software ecosystem two years before this hardware launch, giving content developers, tools makers and engine providers early access through Arm Neural Technology and an open development kit. The now-available Arm Neural Graphics Development Kit includes NSS and NFRU, machine learning extensions for Vulkan, plug-ins for leading game engines including Unreal Engine, and an SDK for custom engines, plus profiling, training and model-optimization tools. The stated goal is a path from experimentation to production without studios building bespoke research infrastructure or abandoning existing workflows.

Tencent Games and Unity China are cited as partners bringing the techniques into familiar development tools. Keli Zhou, engine lead for Where Winds Meet, said the studio is integrating Arm Neural Technology into its Messiah engine and that the game "will be among the first games to bring Arm Neural Technology to players, marking an important step from engine-level integration to real-world deployment."

Why it matters

Arm says more than 14 billion Mali GPUs have shipped to date, so architectural changes here tend to reach enormous volumes of Android-class devices. The approach mirrors the AI-upscaling shift already standard on desktop GPUs from NVIDIA and Intel, but mobile has the harder constraint: efficiency, not raw throughput, is the limiting factor, and Arm is betting that AI-generated pixels are cheaper than rendered ones.

Two caveats are worth holding onto. All the performance figures come from Arm's own benchmarks and demos, and the announcement does not say when the first devices carrying the G2-Ultra NX will reach consumers. Independent numbers will only appear once silicon ships — but if the claims hold even partially, neural rendering moves from a desktop nicety to a mobile baseline.