Arm has shipped its first “AI-native” mobile GPU, and the graphics playbook it borrowed looks a lot like Nvidia’s. The Mali G2-Ultra NX, unveiled September 8, 2026, bakes dedicated neural accelerators into every shader core, enabling AI upscaling, AI frame generation, and AI-assisted ray-tracing denoising on a phone chip for the first time, according to The Register. The technology debuts in a single device: the Xiaomi 18 Fold, a China-only foldable running a custom Xring O3 chipset, as reported by The Verge.
The launch matters beyond one foldable phone. Arm licenses its GPU designs to chipmakers across the Android ecosystem, and this generation is the clearest signal yet that the company wants neural rendering, not just raw shader throughput, to define the next wave of mobile graphics. For the millions of Android devices that don’t run a flagship Qualcomm Snapdragon chip, Arm’s move could close a gap that Nvidia’s DLSS opened on PC and console years ago.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What Is the Arm Mali G2-Ultra NX GPU
The Mali G2-Ultra NX is described by Arm as its first AI-native Mali GPU, built around a new class of hardware block the company calls “NX” — short for neural accelerator. Rather than treating AI graphics as a software layer bolted onto a traditional rendering pipeline, Arm has wired NPU-class accelerators directly into each shader core, according to a deep-dive from Android Authority. That structural choice mirrors how Nvidia embeds Tensor Cores alongside CUDA cores on GeForce RTX GPUs, giving both companies a dedicated engine for running neural network inference without stealing cycles from traditional rasterization work.
Arm frames the NX block as optional and lightweight rather than a power hog. Android Authority’s reporting on the architecture describes NX as “a tightly integrated, optional hardware accelerator that sits inside the GPU’s shader core to run neural graphics tasks while costing as little as 1W in power,” a framing that matters for a device category where battery life and thermal headroom are the real constraints, unlike a desktop GPU with a dedicated cooling loop and a wall socket. The Mali G2-Ultra NX also delivers a roughly 14% conventional rasterization improvement over the prior Mali generation, separate from the neural graphics gains, according to 9to5Google.
The GPU ships as part of a broader platform update that industry coverage has referred to as Arm’s next-generation CSS for Mobile design, pairing the new graphics core with updated CPU clusters aimed at what Arm calls “desktop-class games” and on-device agentic AI workloads, according to 9to5Google’s report on the launch.
Neural Super Sampling: Arm’s Answer to DLSS
The centerpiece of Arm’s pitch is Neural Super Sampling (NSS), an AI upscaling technique that renders a game at a reduced internal resolution and then reconstructs a sharper image using a trained neural network. Arm’s own product materials describe the feature directly: “Neural Super Sampling (NSS) reconstructs a higher-resolution image from lower-resolution rendered inputs, reducing the number of pixels the GPU needs to render natively,” according to Arm’s official Mali G2-Ultra NX product page.
In practical terms, Arm has repeatedly cited a 540p-to-1080p upscale at roughly 4 milliseconds per frame, a figure that first surfaced when the company unveiled NSS at SIGGRAPH 2025 and has carried through into this year’s hardware launch, according to Arm’s newsroom announcement. That’s the same fundamental trade Nvidia makes with DLSS Super Resolution on PC: render less, then use a neural network to fill in detail the eye reads as native quality. The difference is the power envelope. A desktop RTX GPU can draw 300 watts or more under load; a phone SoC has to do this work inside a battery-and-thermal budget measured in single-digit watts.
Arm says the payoff of freeing up those pixels isn’t just a frame-rate number — it’s headroom that developers can spend on lighting, shadow quality, or simply on battery life, since the GPU can hit the same visual target while doing less native rendering work.
Neural Frame Rate Upscaling and Ray-Tracing Denoising
Alongside NSS, Arm is shipping two companion technologies. Neural Frame Rate Upscaling (NFRU) is Arm’s frame-generation feature, generating intermediate frames between two natively rendered ones to raise perceived frame rate — conceptually identical to what Nvidia calls DLSS Frame Generation on PC. Arm’s product page states that “Neural Frame Rate Upscaling (NFRU) uses rendered frames, motion data, and hardware-accelerated optical flow to generate high-quality intermediate frames,” per the company’s official description.
The third piece, Neural Super Sampling and Denoising (NSSD), targets ray-traced scenes specifically. It combines the same upscaling approach as NSS with an AI denoising pass that cleans up the visual noise inherent to ray tracing at low sample counts, letting a phone GPU approximate ray-traced lighting and reflections without the ray budget a console or PC GPU would need. Coverage from GagAdget frames this trio — upscaling, frame generation, and ray reconstruction — as taking “several pages from Nvidia’s” playbook, and the comparison is hard to avoid once you line up the feature names side by side.
Arm summarizes the combined effect plainly: “Together, they reduce graphics workload while enabling higher visual quality and smoother gameplay,” according to the company’s own Mali G2-Ultra NX materials.
Xiaomi 18 Fold: The Launch Device
Every one of Arm’s neural graphics claims currently rests on a single shipping product: the Xiaomi 18 Fold. The foldable runs on a custom Xring O3 chipset that pairs Arm’s newest CPU cores with the Mali G2-Ultra NX GPU, according to reporting from Gizmodo and The Verge. As of the September 8 launch coverage, the phone is sold exclusively in mainland China, meaning the rest of the world is currently reading about desktop-class mobile graphics it cannot yet buy.
Why China First
Xiaomi’s decision to debut a custom silicon platform domestically before a global rollout tracks with a broader pattern among Chinese OEMs building in-house or co-developed chipsets. It also gives Arm a controlled showcase: one flagship device, one well-documented chip, and a foldable form factor that benefits disproportionately from AI upscaling, since folding phones often run games at unusual aspect ratios and resolutions that stress a GPU harder than a standard slab phone.
What’s Missing From the Spec Sheet
Coverage of the launch has focused almost entirely on the GPU and its neural features rather than a full spec dump. Display size, RAM configuration, battery capacity, and pricing for the Xiaomi 18 Fold were not detailed in the outlets that broke the story, including The Verge, Gizmodo, and 9to5Google — a gap that’s worth noting rather than filling in with a guess.
The Performance Numbers Arm Is Publishing
Arm’s most eye-catching performance claim comes from an internal demo called Neural Dawn. According to Android Authority’s deep dive into the architecture, “Arm’s Neural Super Sampling can take Arm’s Neural Dawn demo from an unplayable 15fps to a sustained 60fps with a power budget of just over 2W.” That’s a 4x frame-rate jump attributed almost entirely to the neural pipeline, though it’s worth treating an in-house demo the way you’d treat any vendor benchmark: directionally useful, not a guarantee of what a third-party game will deliver.
| Arm Neural Feature | What It Does | Cited Performance Claim | Closest Nvidia Analog |
|---|---|---|---|
| Neural Super Sampling (NSS) | Upscales a lower-resolution render to a higher output resolution using a neural network | 540p to 1080p at ~4ms per frame | DLSS Super Resolution |
| Neural Frame Rate Upscaling (NFRU) | Generates intermediate frames using motion data and optical flow | Contributes to 15fps → 60fps in Arm’s Neural Dawn demo | DLSS Frame Generation |
| Neural Super Sampling and Denoising (NSSD) | Combines upscaling with AI denoising for ray-traced scenes | Not independently benchmarked in current coverage | DLSS Ray Reconstruction |
| NX accelerator block | Dedicated per-shader-core hardware that runs the neural models above | As little as 1W of added power draw | Tensor Cores |
What’s notably absent from every report reviewed so far is independent, third-party benchmarking in shipping games. Every figure currently in circulation — the 4ms upscale cost, the 15-to-60fps jump, the 14% rasterization gain — originates from Arm’s own materials or from outlets relaying Arm’s demo results. That’s normal for a same-day hardware launch, but it also means the real test comes once reviewers get a Xiaomi 18 Fold and a stopwatch.
How This Stacks Up Against Qualcomm’s Adreno GPUs
Qualcomm’s Snapdragon line, built around its in-house Adreno GPU, has held a performance and marketing edge in mobile gaming for years, particularly among flagship Android phones sold outside China. Arm doesn’t build phones or chips itself — it licenses GPU and CPU designs to companies like Xiaomi, MediaTek, and (for some Exynos generations) Samsung — so its competitive position depends on how attractive its IP is relative to what Qualcomm offers in-house.
NotebookCheck’s analysis frames NSS explicitly as “enhanced upscaling for non-Snapdragon phones,” and notes that Arm’s approach is a step up from Accuracy Super Resolution (ASR), a lighter-weight upscaler Arm introduced earlier. The subtext is straightforward: Arm needs its GPU licensees to have a credible answer to whatever upscaling story Qualcomm tells about Adreno, or those OEMs have less reason to pick Arm’s graphics IP over a fully custom or Qualcomm-supplied alternative. Qualcomm’s own AI PC push, seen recently in devices like the HUMAIN Horizon Ultra built on Snapdragon X2 Elite silicon, shows how central AI acceleration has already become to Qualcomm’s pitch across both phones and PCs.
How This Stacks Up Against Samsung Exynos and AMD RDNA Mobile Graphics
The other comparison point is Samsung’s Exynos line, which uses AMD RDNA-based Xclipse GPUs and already supports hardware ray tracing on recent chips. Arm’s pitch with NSSD is different in emphasis: rather than compete purely on ray-tracing throughput, Arm is betting that AI-assisted reconstruction — rendering fewer rays and using a neural denoiser to clean up the result — can match perceived visual quality while spending less power and silicon than a larger ray-tracing pipeline would need.
Whether that bet pays off depends on execution details none of the current coverage has independently verified yet, including how NSSD holds up in fast-moving scenes or how visible artifacting is compared with native ray tracing on Xclipse-equipped Exynos hardware.
| Platform | GPU Architecture | AI Upscaling Approach | Ray Tracing Support |
|---|---|---|---|
| Xiaomi 18 Fold (Xring O3) | Arm Mali G2-Ultra NX | NSS / NFRU / NSSD (neural, hardware NX block) | Via NSSD AI denoising |
| Qualcomm Snapdragon (Adreno) | Adreno GPU | Qualcomm’s own upscaling and Game Super Resolution features | Hardware-supported on recent flagship Snapdragon chips |
| Samsung Exynos (Xclipse) | AMD RDNA-based Xclipse | No neural upscaling equivalent detailed in current coverage | Native hardware ray tracing via RDNA |
| Nvidia GeForce RTX (PC) | Ada Lovelace / Blackwell | DLSS Super Resolution, Frame Generation, Ray Reconstruction | Dedicated RT cores, mature multi-generation support |
Arm’s Business Case: Licensing, Not Chipmaking
It’s worth remembering what kind of company Arm is in this story. Arm doesn’t manufacture the Xring O3 or sell the Xiaomi 18 Fold — it licenses the Mali G2-Ultra NX design to chipmakers, who then integrate it into their own silicon. That business model means Arm’s win condition isn’t unit sales of one phone; it’s whether enough OEMs and chip designers across the Android ecosystem choose to license this GPU generation over a competing design, whether that’s Qualcomm’s in-house Adreno, AMD’s RDNA for Samsung, or a custom in-house block from another vendor.
Mali and Immortalis GPUs already power a large share of non-Snapdragon Android devices, spanning MediaTek chipsets, some Exynos generations, and various Chinese OEM designs, according to reporting compiled by NotebookCheck and Android Authority. Building neural acceleration into every shader core is Arm’s attempt to keep that footprint relevant as mobile gaming graphics expectations rise, rather than cede the “AI graphics” narrative entirely to Qualcomm and Nvidia’s PC and console ecosystem. It’s a similar dynamic to the one playing out in discrete GPUs, where unverified specs for Nvidia’s next mid-range desktop part have already stirred debate — see the RTX 5070 GB203 core-count rumor — as every GPU vendor tries to stay ahead of the next generation’s expectations before it even ships.
Historical Context: From SIGGRAPH 2025 to a Shipping Phone
This launch didn’t come out of nowhere. Arm first showed Neural Super Sampling at SIGGRAPH 2025, describing it then as its first AI-driven upscaling technology built for mobile GPUs and the debut application of what it called “Arm neural technology,” according to HotHardware’s coverage from that event. At the time, Arm signaled that hardware supporting the feature wouldn’t arrive until sometime in 2026 — a timeline the company has now delivered on, roughly 13 months after the initial reveal.
That gap between announcement and shipping hardware is short by chip industry standards, and it suggests Arm treated NSS as a flagship priority rather than a research curiosity. It also mirrors Nvidia’s own DLSS trajectory: the PC upscaling technology took several generations of GeForce RTX hardware and multiple DLSS versions to mature from a niche feature into something most PC gamers now expect by default. Arm is compressing a similar arc into mobile, but starting from a single launch device rather than a full product stack.
What Analysts and Reviewers Are Saying
Reaction to the launch has been more curious than celebratory. Gizmodo’s write-up on the announcement, titled around the question of whether the technology actually matters, raises a fair point: the real-world impact depends heavily on developer adoption, how many games actually implement NSS and NFRU, and whether phone gamers embrace AI-generated “fake frames” the same way PC gamers have gradually warmed to DLSS Frame Generation after initial skepticism.
Android Authority’s framing leans more optimistic, describing the shift as one that could “transform Android gaming forever” — specifically for the large population of Android users on non-Snapdragon hardware who have historically gotten a lesser graphics feature set than Snapdragon flagship buyers. Both takes can be true at once: the architecture is a genuine step change, and its actual impact on how games look and play won’t be clear until it ships in more than one phone.
Market Impact: Who Gains and Who’s Under Pressure
The most immediate pressure lands on Qualcomm. If Arm’s neural graphics stack proves out in review benchmarks once the Mali G2-Ultra NX reaches more devices, Qualcomm’s Adreno GPUs lose one of their clearer differentiators against cheaper, Arm-licensed alternatives — a dynamic NotebookCheck flagged directly by describing NSS as built for “non-Snapdragon phones.” Samsung’s Exynos/Xclipse team faces a different kind of pressure: their ray-tracing hardware lead looks less unique if Arm’s AI-denoised approach delivers comparable visual results at a lower cost.
For Nvidia, the read is more indirect. Nvidia doesn’t sell mobile GPUs into phones, so there’s no direct product competition. But Arm’s move validates Nvidia’s DLSS architecture as the template every other GPU vendor now has to match, extending Nvidia’s influence over graphics technology even into markets where it doesn’t compete for silicon sales. On the PC side, Nvidia’s own rollout of DLSS 5 to older RTX 40 GPUs and the mixed real-world results seen in NBA 2K27’s DLSS 5 benchmarks are a preview of the scrutiny Arm’s own neural graphics claims will face once independent reviewers get hold of a Mali G2-Ultra NX device.
For Xiaomi, the calculus is about differentiation in a saturated foldable market. Being first — even in a China-only launch — gives Xiaomi a technical story that Samsung, Honor, and Huawei’s competing foldables can’t currently match, at least until their own chip partners license comparable GPU IP.
What This Means for Android Game Developers
New rendering hardware is only useful if developers build for it, and mobile game studios have historically been slower than PC studios to chase bleeding-edge graphics features, since the install base for any single new capability starts at zero. Coverage of the Mali G2-Ultra NX launch notes support for advanced lighting features including Unreal Engine 5’s MegaLights system, which suggests Arm is working to get NSS, NFRU, and NSSD integrated at the engine level rather than asking individual studios to hand-implement support.
That engine-level path is the same one Nvidia took with DLSS: broad Unreal Engine and Unity plugin support did more to spread DLSS adoption across PC games than any single high-profile title did on its own. If Arm can replicate that integration strategy for mobile, the feature could show up in dozens of Android titles within a couple of engine release cycles rather than requiring years of one-off developer partnerships.
Predictions: Where Arm’s Neural Graphics Push Goes Next
- Global expansion in 2027. Coverage from both Android Authority and The Verge points to broader OEM adoption of Arm’s Neural Technology and Mali G2-Ultra NX-class hardware arriving in 2027, once more chip partners tape out designs using the new CSS for Mobile platform.
- Qualcomm responds with its own neural upscaling push. Expect Qualcomm to accelerate its own AI upscaling roadmap for Adreno rather than cede the “non-Snapdragon-only” framing that NotebookCheck’s coverage currently applies to NSS.
- Engine support arrives before broad game-by-game adoption. Given the Unreal Engine 5 MegaLights compatibility already cited in launch coverage, expect Arm to prioritize engine-level plugins over one-off developer deals, similar to Nvidia’s DLSS rollout strategy on PC.
- Independent benchmarks will complicate the marketing numbers. The 15fps-to-60fps and 4ms-per-frame figures currently circulating all come from Arm’s own demos; expect third-party reviewers to publish more conservative real-game numbers once the Xiaomi 18 Fold or a successor device reaches independent testers.
- Foldables become a proving ground for mobile ray tracing. Because foldable displays often render at unusual resolutions and aspect ratios, expect more chipmakers to debut neural graphics features on foldable flagships before rolling them out to mainstream slab phones.
The Limits of What’s Confirmed So Far
It’s worth being precise about what is and isn’t verified at this stage. Pricing for the Mali G2-Ultra NX licensing terms and for the Xiaomi 18 Fold itself has not been disclosed in any of the outlets covering the launch. No named Arm or Xiaomi executive has been quoted directly in the coverage reviewed; statements are attributed to the companies generally or to their official product and newsroom pages. And every performance figure in circulation — the frame-rate jump, the millisecond cost per frame, the power draw — comes from Arm’s own demos rather than from an independent lab. None of that makes the technology less real; it just means the gap between “vendor claim” and “verified real-world result” hasn’t closed yet.
Why This Launch Is Bigger Than One Foldable Phone
Mobile gaming graphics have lagged console and PC for a simple reason: power and thermal budgets that would shut a desktop GPU down within minutes. AI-assisted rendering is the first technique in years that promises to narrow that gap without simply demanding more silicon and more battery. Arm shipping a working implementation — even in a single, region-locked device — puts pressure on every other mobile GPU vendor to show its own answer, and it puts pressure on game engines to treat neural upscaling as a mobile-first feature rather than a PC-only afterthought.
The next twelve months will tell whether Neural Super Sampling becomes as standard on Android as DLSS became on PC, or whether it stays a showcase feature confined to a handful of flagship devices while the broader Android market waits for a cheaper, more power-hungry generation to catch up. For a broader look at how AI acceleration is reshaping chip design across categories, see tech-insider.org’s coverage of AI chips, and for context on how quickly gaming hardware trends shift underneath developers, our Steam Hardware Survey breakdown tracks a similar pattern on the PC side.
Frequently Asked Questions
What is the Arm Mali G2-Ultra NX?
It’s Arm’s first AI-native mobile GPU, announced September 8, 2026, with dedicated neural accelerator hardware (“NX”) built into each shader core to power AI upscaling, frame generation, and ray-tracing denoising, according to The Register and Arm’s own newsroom.
Which phone has the Mali G2-Ultra NX GPU?
As of the September 2026 launch, only the Xiaomi 18 Fold ships with it, running a custom Xring O3 chipset. The device is currently sold exclusively in mainland China, according to The Verge and Gizmodo.
How is Arm’s Neural Super Sampling different from Nvidia DLSS?
Conceptually they work the same way: both render at a lower resolution and use a neural network to reconstruct a higher-resolution image. The difference is the target hardware — NSS is built for phone-level power budgets (Arm cites roughly 4ms per frame for a 540p-to-1080p upscale), while DLSS runs on desktop and laptop GPUs with far larger power envelopes.
Does the Mali G2-Ultra NX support ray tracing?
It supports AI-assisted ray tracing through a feature called Neural Super Sampling and Denoising (NSSD), which combines upscaling with a neural denoiser to clean up ray-traced scenes rendered with fewer samples, per Arm’s official product materials.
When will Arm’s neural graphics technology reach phones outside China?
Current coverage from Android Authority and The Verge points to broader global availability arriving in 2027, as more chip partners adopt Arm’s newest mobile platform and license the Mali G2-Ultra NX design.
Is this the same as Qualcomm’s upscaling technology on Snapdragon phones?
No. Arm’s Neural Super Sampling is built into Arm’s own Mali GPU architecture, which Qualcomm doesn’t use — Snapdragon chips run Qualcomm’s in-house Adreno GPU with its own separate upscaling features. NotebookCheck’s coverage specifically describes NSS as targeting “non-Snapdragon phones.”
How much does the Mali G2-Ultra NX GPU cost?
Neither Arm’s licensing pricing nor the retail price of the Xiaomi 18 Fold has been disclosed in coverage of the launch. Arm licenses chip designs to manufacturers rather than selling GPUs directly to consumers.
What performance gains has Arm actually demonstrated?
In an internal demo called Neural Dawn, Arm’s Neural Super Sampling took frame rate from an unplayable 15fps to a sustained 60fps at a power budget of roughly 2W, according to Android Authority’s architecture deep dive. That figure comes from Arm’s own demonstration rather than independent third-party benchmarking.
Related Coverage
- Nvidia Brings DLSS 5 to RTX 40, Skips Full Control [2026]
- Xiaomi 18 Fold: Arm Debuts New GPU China-Only [2026]
- NBA 2K27 DLSS 5 Verdict: 4 of 6 RTX 50 GPUs Proven [2026]
- PlayStation 6 Cost Hits $960 as Memory Crunch Bites [2026]
- NVIDIA RTX 5070 GB203 Rumor: 8,960 Cores Eyed [2026]
