The next large improvement in phone graphics may not come from rendering more pixels. It may come from asking the GPU to calculate fewer of them—and letting dedicated AI hardware reconstruct the rest.
Arm’s new Mali G2-Ultra NX brings neural acceleration directly into a mobile GPU’s shader cores.
The architecture supports AI-assisted resolution upscaling, generated intermediate frames and denoising for ray-traced graphics. These ideas are already familiar to PC players through technologies such as Nvidia DLSS, AMD FSR and Intel XeSS.
Bringing them into mainstream mobile silicon could help demanding games look sharper and run more smoothly without requiring the phone to consume dramatically more power.
The AI hardware sits inside the graphics pipeline
Phone processors already contain neural-processing hardware, but a separate NPU is not always the most efficient place to execute graphics work.
Mali G2-Ultra NX integrates its neural accelerators inside the GPU’s shader cores. They share the graphics processor’s memory system, caches and control structures.
That tighter connection reduces the need to move large quantities of frame data between different parts of the chip.
The accelerators support INT8 and INT16 calculations and include hardware for optical-flow analysis. Optical flow helps the system estimate how objects move between frames, which is particularly important when generating a new frame that the game engine never rendered conventionally.
This is not a chatbot bolted onto a graphics chip. It is specialized machine-learning hardware designed to participate directly in rendering.
Arm has three different neural graphics tools
Neural Super Sampling renders a game at a lower internal resolution and reconstructs a higher-resolution output using current-frame information, motion data and previous frames.
The principle is simple: calculating every pixel at full resolution is expensive. Reconstructing a credible final image can leave more graphics capacity available for lighting, geometry or higher frame rates.
Neural Frame Rate Upscaling generates intermediate frames between conventionally rendered ones. A game producing 30 real frames per second can display additional AI-generated frames to create smoother visible motion.
Neural Super Sampling and Denoising combines image reconstruction with the removal of noise from ray-traced lighting. Ray tracing often uses a limited number of expensive samples, leaving imperfect data that must be cleaned before presentation.
Together, these techniques allow the GPU to perform less traditional rendering work while still presenting a detailed final image.
The fourfold performance claim needs translation
Arm’s Neural Dawn technology demonstration reportedly moves from approximately 15fps to 60fps when its neural techniques are combined.
Arm describes this as up to four times greater performance efficiency, accompanied by as much as 70% less external-memory traffic compared with native rendering.
That does not mean G2-Ultra NX contains four times the conventional graphics performance of its predecessor.
The demonstration begins with a lower-resolution image, reconstructs it from 540p to 1080p, cleans its ray-traced lighting and inserts generated frames between the frames the GPU actually renders.
Arm’s more conventional comparison is considerably smaller: up to 14% better non-AI gaming performance than its previous flagship configuration. Depending on the workload, independent technical analysis places expected rasterisation gains in roughly the 9% to 20% range.
The neural features are therefore the major architectural change. Raw generation-to-generation performance is useful but less dramatic.
Generated frames improve smoothness, not responsiveness
A displayed frame and an interactively rendered frame are not equivalent.
Frame generation can make camera movement and animation look smoother. The generated frame does not contain a newly processed controller or touchscreen input, because it is created between two conventionally rendered frames.
A game visibly presenting 60fps through frame generation can therefore retain input latency closer to its lower underlying frame rate.
The system also needs enough information to predict movement correctly. Fast effects, interface elements, particles and objects suddenly entering the scene can produce ghosting, distortion or unstable detail.
These problems have improved on PCs, but they have not disappeared. Phone screens are smaller, which may conceal some imperfections, while competitive games make additional latency particularly noticeable.
The GPU receives meaningful conventional upgrades too
G2-Ultra NX is not entirely dependent on AI reconstruction.
Arm has redesigned its execution engine in what it calls Mali’s largest instruction-set change in seven generations. The GPU supports more registers per processing group, helping it handle larger and more complicated shaders without repeatedly reaching into slower external memory.
The third-generation ray-tracing unit uses more compact geometry representations and introduces hardware support for opacity micromaps.
Opacity micromaps help the GPU process detailed foliage, fabrics and other partially transparent surfaces without performing the most expensive calculation for every ray.
Arm claims 13% less DRAM traffic in selected ray-tracing benchmarks. One internal demonstration gained 30% in frame rate while reducing its ray-tracing workload by as much as 70%.
These remain manufacturer benchmark results rather than a guarantee for finished retail games.
Games must be designed to use it
Existing games will not automatically gain every neural feature when they run on a new phone.
Developers need to provide motion vectors, depth information and other engine data. They must integrate Arm’s tools, evaluate image quality and decide whether the latency and reconstruction tradeoffs suit their game.
Arm supplies a Neural Graphics Development Kit, Vulkan machine-learning support and integrations for engines including Unreal Engine. That reduces the engineering burden, but it does not turn neural rendering into a universal settings switch.
Where Winds Meet, Infinity Nikki and Arena Breakout: Infinite are among the first announced games supporting Arm’s neural graphics technology.
Initial implementations will concentrate largely on super sampling. Arm told The Verge that frame generation is expected in early 2027, with its more advanced denoising technology following later.
Xiaomi gets the hardware first
The Xiaomi 18 Fold is the first announced phone containing Mali G2-Ultra NX through Xiaomi’s custom XRing O3 processor.
It launched in mainland China on September 7. Xiaomi’s large Pad 9 Pro Max tablet also uses the chip, giving the GPU a larger battery and more surface area for cooling.
Arm says additional phones and larger-screen devices should follow in 2027. Neural graphics will also extend beyond the highest Ultra configuration into lower Arm GPU tiers.
That broader adoption matters more than one expensive Chinese foldable. Developers will devote serious resources to these features only if enough compatible devices reach players.
What players may actually notice
The best outcome is not necessarily console-quality graphics squeezed onto a phone.
A more practical benefit would be stable 60fps gameplay at a sharp-looking resolution without severe thermal throttling after 20 minutes.
Developers could also use the saved power to extend battery life rather than raising visual quality. That may be more valuable to average players than extra ray-traced reflections.
Simple puzzle, strategy and casual games will gain little from this architecture. It is primarily relevant to graphically demanding 3D titles and future ports of PC or console games.
No one should buy a phone merely because its specification sheet lists neural graphics. Buyers need to know which games support the feature, how the image holds up in motion and whether performance remains stable during a long session.
Real Talk: reconstructed pixels still need scrutiny
AI-assisted rendering is not free performance. It exchanges native image data, latency and occasional visual errors for efficiency.
That trade can be excellent. DLSS and competing PC technologies have shown that well-trained reconstruction can produce an image that looks better than conventional low-resolution rendering while consuming far less graphics power.
Poor implementations can also create shimmering edges, broken interface elements and motion artifacts. Frame generation can inflate the displayed frame-rate number without making controls equally responsive.
Arm’s fourfold demonstration represents a best-case combination of several technologies inside a controlled workload. Retail games, different chip configurations and thin phone cooling systems may produce less dramatic results.
The hardware is credible. The experience still depends on developers using it carefully.
IskraCore Take
Mali G2-Ultra NX is more important than a routine mobile GPU speed increase because Arm has placed dedicated neural acceleration directly where games can use it.
Upscaling and denoising attack the real constraints of mobile graphics: heat, battery capacity and limited memory bandwidth.
The fourfold headline should not be confused with four times more raw GPU power. Part of that gain comes from reconstructed pixels and generated frames, with familiar compromises in latency and image stability.
Success will be measured through supported games and sustained performance—not technology demonstrations or theoretical frame counters.
Phones do not need to render like desktop GPUs. They need to become smarter about which work is worth rendering at all.