TL;DR: Arm's Mali G2-Ultra NX puts dedicated AI accelerators inside a mobile GPU to upscale resolution and generate frames, debuting in China's Xiaomi 18 Fold on Sept. 8, 2026.

Arm introduced a new mobile graphics system on Sept. 8, 2026, that uses on-chip artificial intelligence to raise frame rates and image quality without raising power use, and the first phone to carry it launched the same day in mainland China, according to The Verge.

The debut device is the Xiaomi 18 Fold, a book-style foldable powered by Xiaomi's custom Xring O3 chip, which integrates the Arm Mali G2-Ultra NX graphics processor, according to The Verge. Arm described the launch as part of its CSS for Mobile 2 compute platform, which combines the new GPU with a new C2 CPU cluster and system IP for agentic AI and AI-native graphics, according to the Arm Newsroom.

Arm's Chris Bergey, executive vice president for Edge AI, told The Verge in an interview that the technology follows five years of development and brings an AI gaming graphics accelerator inside the GPU shader. Bergey said shoppers outside China will be able to buy devices with the graphics technology next year, including phones and larger-screen devices such as tablets or Chromebooks, according to The Verge.

What launched in China on Sept. 8

The Xiaomi 18 Fold is Xiaomi's first book-style foldable since the Mix Fold 4 from 2024, and it adopts a short, wide design with a 7.58-inch inner display and a 5.38-inch outer display, according to The Verge's separate hands-on report on the device.

Xiaomi positioned the 18 Fold around its Xring O3 chipset, which Xiaomi says delivers Snapdragon 8 Elite-level performance and introduces the Mali G2-Ultra NX GPU with mobile-first support for Arm's DLSS-style acceleration on Android, according to The Verge. The phone includes a 6,000mAh silicon-carbon battery, support for 67W wired charging and 50W wireless charging, dual OLED displays with up to 4,000 nits of peak brightness, and a Leica-partnered triple rear camera with a 200-megapixel main camera, a 50-megapixel 3.5x periscope telephoto and a 50-megapixel ultrawide, according to The Verge.

The 18 Fold was available to preorder in China from ¥10,999, about $1,600, with official sales scheduled for Sept. 10, while Xiaomi had not confirmed whether the phone will launch internationally, according to The Verge. Xiaomi also showed the Xiaomi Pad 9 Pro Max tablet with the same Xring O3 chip, which Xiaomi claims can play games at 3.2K resolution and 120 frames per second, according to The Verge.

What Mali G2-Ultra NX adds to a phone GPU

Arm calls the new capability neural graphics, a set of hardware-accelerated techniques that reconstruct, enhance and generate visual detail rather than rendering every pixel and frame in the traditional way, according to the Arm Newsroom.

The three named techniques are Neural Super Sampling, which reconstructs a higher-resolution image from a lower-resolution render, Neural Frame Rate Upscaling, which generates intermediate frames to increase frame rates, and Neural Super Sampling and Denoising, which combines upscaling with denoising for ray-traced scenes with complex lighting and shadows, according to the Arm Newsroom technical blog.

Arm said the NX designation refers to dedicated neural accelerators tightly integrated into the shader cores, allowing neural graphics workloads to run alongside graphics and compute workloads while reusing the GPU memory system, coherent caches and control structures to reduce data movement, according to the Arm Newsroom.

The neural accelerator supports INT8 and INT16 operations, includes its own motion-engine processing block, can operate at up to twice the main GPU clock at the same voltage, and shares memory with the rest of the GPU subsystem to maintain memory coherency, according to Android Authority. Arm said the tight integration keeps the added power cost as low as about 1W in some configurations, according to Android Authority.

How Arm describes the performance change

Arm separated traditional rasterization gains from gains when neural techniques are active.

For existing game content without AI assistance, Arm reported up to 14% higher performance compared with the previous generation, according to the Arm Newsroom. In reference-platform comparisons cited by Android Authority, gaming performance increased by 9% in Honkai Star Rail to 14% in Fortnite when comparing a 14-core Mali G1-Ultra reference platform with a G2-Ultra equivalent, with differences in shader-core clock at 1.53GHz versus 1.70GHz, and Arm expects 17% to 20% improvements in demanding mobile graphics benchmarks at the upper end, according to Android Authority.

The Verge reported an additional comparison tied to Xiaomi's prior hardware, describing an alleged 85% performance increase over the GPU Xiaomi was using previously, alongside the 14% increase over Arm's previous flagship in non-AI gaming, according to The Verge.

With neural graphics active, Arm claims up to four times the frame rate when Neural Super Sampling, Neural Frame Rate Upscaling and Neural Denoising are applied together, because each step reduces conventional rendering work and relies on AI to fill in remaining pixels and frames, according to The Verge. Arm also said the combination delivers up to 4x higher performance per watt for neural graphics, according to the Arm Newsroom.

In the Neural Dawn demonstration developed with Sumo Digital, Arm said NX can lift output from 15 frames per second to a sustained 60 frames per second with a power budget of just over 2W, paired with a 70% reduction in DRAM memory traffic compared with the prior G1-Ultra graphics core, according to Android Authority. The demonstration path described uses 2x Neural Super Sampling and Denoising upscaling from 540p to 1080p, including ray-tracing denoise in the same pass, followed by Neural Frame Rate Upscaling to double the frame rate from 30 fps to 60 fps, according to Android Authority.

Arm also linked the efficiency gain to longer sessions. Bergey told The Verge a phone could play at 1080p and 60 frames per second for three hours in an example scenario, compared with 30 minutes in the past, according to The Verge. Bergey said an entire phone with these chips will consume no more than 2.5 watts in that type of sustained-gaming setup, with 1.5 watts allotted to the GPU, according to The Verge.

Why power and heat matter for sustained play

Mobile sustained performance is limited by power, heat and memory bandwidth, and Arm framed the architecture as a way to hold higher frame rates without making a phone get warm or demand a charger quickly, according to The Verge.

Bergey said device and chipset makers can either use the headroom for higher resolution and frame rates or run the GPU at lower wattage to extend battery life, adding that the approach is intended as mainstream rather than overclocking, according to The Verge. With the Xiaomi Pad 9 Pro Max, which pairs the GPU with a larger battery, larger screen and more surface area for cooling, Arm said the chip may offer additional performance, with the Chinese company showing 3.2K at 120 frames per second, according to The Verge.

The underlying GPU changes extend beyond AI. Arm introduced a new execution engine that it described as its biggest Mali GPU instruction-set architecture upgrade in seven generations, providing up to twice the registers per warp to handle demanding scenes more efficiently, according to the Arm Newsroom technical blog. Android Authority reported the register file grows from 64 to 128 registers per warp with more dynamic allocation, supporting 16 warp-wide steps from 16 to 128 registers, a change intended to let advanced shader technologies such as Unreal Engine 5's Lumen and Nanite run without register spilling and associated performance drops from reaching DRAM, according to Android Authority.

For ray tracing, Arm introduced its third-generation ray-tracing unit with a more compact triangle representation using shared edges to reduce duplication and fit more geometry in cache, lowering DRAM traffic by 13% on leading ray-tracing benchmarks in Arm's description, according to Android Authority. The design also adds hardware-accelerated opacity micromaps for the first time, aimed at transparent geometry such as foliage, fabrics and layered surfaces, with combined reference-design gains of about 18% to 24% better ray-tracing performance than the predecessor depending on benchmark, according to Android Authority. In one Arm demo cited by the Arm technical blog, opacity micromaps increased frame rates by 30% and reduced ray-tracing workload by up to 70% in that scene.

What developers must do before players see it

The new upscaling and frame-generation features are not a system-wide switch that accelerates every installed game.

Like the most advanced versions of Nvidia, AMD and Intel techniques, developers will need to integrate Arm's features into individual games, although participating engines including Unreal Engine and Unity can help carry the integration, according to The Verge. Bergey said he does not expect gamers to be able to enable the features for any game they choose, according to The Verge.

Arm said leading engines and studios are already integrating Arm Neural Technology into developer flows, including Tencent Games Central Tech's MagicDawn and Unity China's Tuanjie Engine, according to the Arm Newsroom. For specific titles, Arm said NetEase's Where Winds Meet is planning to bring its NSS-enabled version to players this year, alongside an NSSD technology demonstration using Tencent Games' Arena Breakout Infinite and NSS integration with Infold Games' Infinity Nikki, according to the Arm Newsroom.

The rollout order is staggered. Bergey said the first titles will largely use super sampling, with frame generation planned for early 2027 and denoising planned soon after that, according to The Verge. Android Authority similarly reported that NSS was published last year while frame-rate upscaling and related models remained in different stages of development at the time of the announcement.

Arm also pointed to its Neural Graphics Development Kit, KleidiAI software work and a new AI Portal as the developer entry point for optimized models, code and tools, saying new hardware matters only when developers can reach it through familiar frameworks, according to the Arm Newsroom.

Where the technology goes after Xiaomi

Arm said the Mali G2-Ultra NX targets flagship smartphones, while neural graphics will also be available across Premium and Pro configurations to reach broader mobile segments and tiers, according to The Verge.

Bergey declined to name other devices using Mali G2-Ultra NX but suggested that some upcoming Googlebooks might use the graphics technology, according to The Verge. He also said buyers will be able to get the technology in phones next year and in larger-screen devices including tablets or Chromebooks, according to The Verge.

Arm placed the GPU inside the wider CSS for Mobile 2 platform, which adds the C2 CPU cluster with C2-Ultra and C2-Pro CPUs and two SME2 units for responsive on-device AI, according to the Arm Newsroom. Compared with C1-Ultra, Arm said C2-Ultra delivers up to 1.7x higher AI performance and 15% higher single-thread performance while using up to 38% less power at the same performance level, with doubled SME2 capability enabling a 70% speedup on recent small language models, according to the Arm Newsroom.

Arm also cited scale as a distribution advantage, noting more than 14 billion Mali GPUs shipped to date, according to the Arm Newsroom technical blog. The company argued that scale and common platform software could help neural graphics move from demonstrations into production games faster than a standalone GPU feature, according to the Arm Newsroom.

Sources referenced