Arm used its Arm Everywhere China event on September 8, 2026 to unveil the second generation of its mobile compute subsystem, and the headline pairing inside it is significant: a new C2-Ultra CPU cluster and the Mali G2-Ultra NX GPU, the company’s first graphics processor built with neural accelerators wired directly into its shader cores. Arm calls the combined package CSS for Mobile 2, and according to Arm’s own newsroom post, it folds the new CPU cores, the new GPU, an updated system interconnect, and a refreshed software stack into one licensable platform for phone makers.
The announcement matters because it sets the baseline silicon that Android flagship phones will ship with in 2027. Arm doesn’t build phones or chips itself, it licenses processor designs to companies like MediaTek, Qualcomm, and Samsung, who then build actual system-on-chip products around that IP. When Arm changes the CPU and GPU blueprint, the effects show up a year or so later in Geekbench listings, gaming benchmarks, and eventually the phone in your pocket. It’s the latest move in a broader race across the industry’s AI chip landscape, where CPU and GPU makers are all chasing faster, more efficient on-device inference.
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 Arm Actually Announced on September 8
CSS for Mobile 2 is Arm’s second-generation Compute Subsystem for Mobile, a pre-integrated bundle of CPU, GPU, and interconnect IP that chip designers can adopt as a starting point rather than stitching components together individually. In its September 2026 coverage, Notebookcheck reported that the new platform “combines the new Arm C2-Ultra and C2-Pro CPU cores with Mali G2-Ultra NX graphics, an updated system interconnect and a new software stack,” and that same month’s Notebookcheck GPU database entries confirm the Mali G2-Ultra NX GPU as the graphics component anchoring the bundle. That interconnect, referred to as SI L2 in Arm’s materials, is meant to move data between the CPU cluster, the GPU, and shared memory with less overhead than the previous generation.
The framing Arm chose for the launch leans heavily on “agentic AI,” the idea that phones will increasingly run AI agents locally rather than routing every request to the cloud. Arm’s newsroom said the platform is built for this shift, pairing CPU-side matrix acceleration with a GPU that runs neural network operations inside the same pipeline as traditional 3D rendering. Whether that framing translates into consumer-visible features on day one is unclear, since most of what’s confirmed so far describes the underlying architecture rather than shipping apps.
C2-Ultra and C2-Pro: The CPU Side of the Story
The Arm C2 CPU cluster mixes two types of cores: the performance-oriented C2-Ultra, which Arm describes as its most powerful mobile CPU core to date, and the efficiency-focused C2-Pro, aimed at background and lighter workloads. Both are paired with two SME2 (Scalable Matrix Extension 2) units built directly into the CPU cluster, hardware dedicated to the matrix math that machine learning inference relies on.
Arm’s reported performance figures put the C2 generation at up to 15% higher single-thread performance and up to 12% higher multi-thread performance compared with last year’s C1 cores. On AI workloads specifically, using both SME2 units together, Arm claims up to 1.7x faster performance on current AI models while consuming 38% less power than the C1-based approach. That power figure is the more interesting number for phone makers, since sustained AI inference on a handset is thermally constrained in ways that a fan-cooled laptop chip is not.
None of the outlets that covered the launch, including Notebookcheck and CNET, published detailed micro-architecture specifics like cache sizes or pipeline depth for C2-Ultra. What’s confirmed is functional: two SME2 blocks, a performance/efficiency core split, and the percentage gains above. Readers should treat anything beyond those figures, including specific clock speed targets for individual phone SoCs, as speculation until a chip vendor like MediaTek or Qualcomm confirms its own implementation.
Mali G2-Ultra NX: Arm’s First AI-Native GPU
The GPU side is where Arm is making its boldest claim. Mali G2-Ultra NX, which Arm officially launched on September 8, 2026, is described across Arm’s own materials and independent outlets as the company’s first “AI-native” Mali GPU, meaning dedicated neural accelerators sit inside the shader cores themselves rather than as a bolt-on neural processing unit elsewhere on the chip. As of that September 2026 launch, Arm says those NX tensor cores are wired directly into every shader core across the GPU, a broader rollout than the partial, half-populated configuration that earlier pre-launch database listings had implied. Those accelerators share the GPU’s memory system, coherent caches, and control logic with the traditional graphics pipeline, which in theory cuts down the data shuffling that happens when a phone splits a workload between a GPU and a separate NPU.
Chips and Cheese, a hardware architecture outlet that dug into Arm’s technical disclosures, described Mali G2-Ultra NX as “the largest re-architecting of the GPU IP in 7 generations.” That’s a notable claim for a company that has iterated on Mali steadily for over a decade. According to that same analysis, a full G2-Ultra NX configuration tops out at 24 shader cores, retains 128 FMA (fused multiply-add) units, and delivers 256 FP32 FLOPs per clock or 512 FP16 FLOPs per clock, roughly in line with the raw throughput ceiling of the prior generation even as the neural accelerator hardware is new.
Ray Tracing Gets a Third Generation
Mali G2-Ultra NX also introduces a new execution engine alongside third-generation ray tracing units, referred to as RTUv3 in Arm’s technical documentation, added to the GPU at its September 2026 launch. This supersedes the RTUv2 hardware from the previous Mali generation. Ray tracing on mobile has been a slow build for Arm; earlier Mali GPUs supported it in limited form, and RTUv3 is positioned as the version mature enough to handle more demanding lighting effects without collapsing frame rates. Arm has now put a number behind that ambition, designing G2-Ultra NX to sustain mobile gaming at up to 120 frames per second, though independent game benchmark numbers from reviewers had not been published as of this writing.
Notebookcheck’s own GPU database, updated in September 2026, lists a specific configuration, the Mali G2-Ultra NX MP16, with 16 compute units, Vulkan 1.4 support, and 4MB of L2 cache, tied to a listed release date of August 22, 2026, roughly two weeks before Arm’s public unveiling. That same September 2026 entry specifies that only 8 of the MP16’s 16 compute units are built with dedicated NX tensor cores, meaning the AI-native hardware is split across half the GPU rather than present uniformly across every unit. That timing, and the partial-unit rollout, suggests some silicon partners had early access to the design before the formal announcement, which is standard practice for Arm’s licensing business.
Who’s Actually Building Chips With This
Arm licenses designs, it doesn’t fabricate chips, so the real test of CSS for Mobile 2 is which silicon vendors adopt it and in what products. The clearest signal now comes from an actual shipping device: Xiaomi launched the 18 Fold on September 8, 2026 with Mali G2-Ultra NX built into its in-house Xring O3 chip, making it the first commercially available phone confirmed to run Arm’s new GPU. Alongside that, a Geekbench listing spotted by Notebookcheck for a Vivo X500 Pro Max showed a GPU entry reading “Mali-G2-Ultra-NX MC12,” a 12-core configuration, tied to a MediaTek Dimensity 9600 Pro chip clocked at 4.55GHz on its prime cores, suggesting MediaTek is close behind Xiaomi in bringing the architecture to market.
Beyond that one listing, CNET and other outlets frame the platform as targeting 2027 Android phones broadly, without naming additional specific OEMs or chip designs. Qualcomm and Samsung, Arm’s two other major mobile licensees, have not been quoted or referenced in connection with this specific launch in the sources reviewed for this piece. That’s a meaningful gap: Qualcomm’s Snapdragon 8 Elite Gen 6 already exists as a separate, competing design path, and it’s not yet clear whether or when Qualcomm’s next flagship chip will incorporate Arm’s new C2 or Mali G2-Ultra NX blocks rather than Qualcomm’s own custom Oryon cores and Adreno graphics.
C2-Ultra and Mali G2-Ultra NX vs. the Previous Generation
The table below summarizes the confirmed generational comparison between the new platform and its predecessor, based on figures Arm and independent outlets have published so far.
| Metric | Previous Generation (C1 / earlier Mali) | New Generation (C2-Ultra / Mali G2-Ultra NX) |
|---|---|---|
| Single-thread CPU performance | Baseline | Up to 15% higher |
| Multi-thread CPU performance | Baseline | Up to 12% higher |
| AI model performance (2x SME2) | Baseline | Up to 1.7x faster |
| Power draw on AI workloads | Baseline | 38% lower |
| Neural accelerators in shader cores | Not present | Dedicated accelerators integrated |
| Ray tracing generation | RTUv2 | RTUv3 (new execution engine) |
| Max shader cores (GPU) | Up to 24 (prior gen ceiling) | Up to 24 (unchanged ceiling) |
| FMA units per core config | 128 | 128 (unchanged) |
| FP32 throughput | 256 FLOPs/clock | 256 FLOPs/clock (unchanged) |
The raw shader math is largely unchanged, which tells you the real story here isn’t more brute-force horsepower, it’s the new neural accelerator hardware layered into the same shader budget, plus the jump to RTUv3 ray tracing. Arm is betting that architecture, not core count, is what moves the needle for on-device AI graphics next.
How This Stacks Up Against Qualcomm, Apple, and MediaTek
Arm’s announcement lands in a mobile chip market where each major player has taken a different bet on where AI processing should live. Qualcomm has invested heavily in its own custom Oryon CPU cores paired with Adreno graphics and a dedicated Hexagon NPU, keeping AI acceleration as a separate block rather than folding it into the GPU shader pipeline the way Arm’s Mali G2-Ultra NX does. Apple’s approach with its A-series and M-series chips similarly keeps a discrete Neural Engine alongside its GPU rather than merging the two.
MediaTek, by contrast, licenses core Arm CPU and GPU IP more directly for many of its Dimensity chips, which is why the Vivo X500 Pro Max Geekbench listing showing a Dimensity 9600 Pro paired with Mali-G2-Ultra-NX graphics is the clearest early signal of adoption. If that pairing holds through to a commercial launch, MediaTek would be the first vendor to ship Arm’s new AI-native GPU architecture in a real product, giving it a talking point against Qualcomm’s Snapdragon 8 Elite Gen 6 line in the same flagship tier. On the discrete GPU side of the industry, AMD’s own RDNA 5 architecture leaks point to a similar industry-wide push toward higher clocks and better ray tracing throughput, though that effort targets desktop and laptop graphics cards rather than mobile SoCs.
The bigger competitive question is whether integrating neural acceleration directly into the shader cores, rather than keeping it as a separate NPU block, actually produces better real-world results for game developers and AI app builders. Arm’s argument is that shared memory and cache reduce latency and power for workloads that blend graphics and AI, and Arm has since detailed what that looks like in practice: as of September 2026, G2-Ultra NX supports neural upscaling, frame-rate upscaling, and denoising directly on the GPU, with Taiwanese outlet TechBang reporting specific figures of 4x super-resolution upscaling and 2x frame generation. Qualcomm and Apple’s separate-block approach has the advantage of letting each unit be optimized and scaled independently. Neither side has published head-to-head benchmarks against the other’s approach, so this remains an architectural debate rather than a settled performance verdict.
Reported Specifications at a Glance
The following table pulls together the specific configuration details that have been published for the Mali G2-Ultra NX variants tracked so far, plus the CPU cluster’s confirmed building blocks.
| Component | Confirmed Detail | Source |
|---|---|---|
| Platform name | Arm CSS for Mobile 2 | Arm Newsroom |
| CPU cores | C2-Ultra (performance) + C2-Pro (efficiency) | Arm Newsroom |
| AI matrix units | Two SME2 units integrated in CPU cluster | Arm Newsroom |
| GPU name | Mali G2-Ultra NX | Arm Newsroom / Notebookcheck |
| Mali G2-Ultra NX MP16 config | 16 compute units (8 with dedicated NX tensor cores), Vulkan 1.4, 4MB L2 cache | Notebookcheck GPU database (Sept. 2026) |
| Mali G2-Ultra NX MC12 config | 12 cores, paired with Dimensity 9600 Pro in Geekbench listing | Notebookcheck (Vivo X500 Pro Max leak) |
| MP16 listed release date | August 22, 2026 | Notebookcheck GPU database |
| Platform announcement date | September 8, 2026, Arm Everywhere China | Arm Newsroom |
| Ray tracing hardware | Third-generation ray tracing units (RTUv3) | Arm technical disclosure |
Why Arm Is Betting on Neural Graphics Now
Arm’s own description of the launch frames it as part of a broader push into what it calls the agentic era of computing, where AI agents run tasks continuously on-device rather than as occasional cloud queries. According to Arm’s official newsroom statement, “The Arm C2 Ultra CPU with SME2 brings a new level of responsive, efficient on-device AI to the CPU, while the Arm Mali G2-Ultra NX GPU introduces dedicated neural accelerators that enable neural graphics to run natively on the GPU.” (Arm Newsroom)
Arm has also positioned the GPU specifically around gaming and visual workloads. On its consumer smartphone product page, Arm states that “Mali G2-Ultra NX GPU through dedicated neural network accelerators, new ray tracing, and sustained stable stronger performance, brings AI-native graphics to mobile gaming.” (Arm product page) That framing puts mobile gaming, rather than productivity apps, at the center of the pitch for why phone makers should adopt this architecture over sticking with a conventional GPU plus separate NPU layout.
Arm’s Chinese-language product materials go further into the technical rationale, describing the GPU as combining dedicated neural network accelerators, a new execution engine, and third-generation ray tracing units to deliver richer, more immersive graphics for mobile games under strict power and thermal constraints, according to Arm’s own product documentation. TechBang’s September 2026 reporting put a specific figure on that constraint, citing a target of just 1 to 2 watts for sustained high-frame-rate gaming on G2-Ultra NX silicon. The repeated emphasis on power and heat limits in Arm’s messaging reflects the core engineering challenge of putting more AI compute into a phone-sized thermal envelope without throttling.
Historical Context: Arm’s Mali Roadmap
Mali has had an uneven reputation compared with Qualcomm’s Adreno and Apple’s in-house GPU designs, often trailing in raw gaming performance despite broad adoption across MediaTek and budget-to-midrange Samsung Exynos chips. Arm has iterated on ray tracing support gradually over recent Mali generations rather than in one leap, and Chips and Cheese’s characterization of G2-Ultra NX as the largest re-architecture in seven generations suggests Arm sees this as a bigger structural break than its recent yearly Mali updates, which mostly added incremental clock and efficiency gains rather than new fixed-function hardware blocks.
On the CPU side, Arm’s SME (Scalable Matrix Extension) technology first appeared in server and desktop-class Arm cores before making its way into mobile designs, and the CSS for Mobile 2 platform marks a deeper mobile-specific integration of SME2 with two units dedicated to matrix operations rather than one. That progression mirrors how Arm has historically trickled server and PC-class features down into phone silicon roughly a generation or two later, once power and die-area budgets allow it. It also echoes the broader CPU industry pattern seen in x86, where chips like AMD’s Ryzen 9950X3D2 and Intel’s upcoming Nova Lake generation keep adding specialized silicon for AI workloads rather than just chasing higher clock speeds.
What Game Developers and App Builders Should Watch
For developers building mobile games or AI-powered apps, the practical question is tooling support. Arm’s platform bundle includes a refreshed software and developer tools stack alongside the CPU and GPU IP, which historically has meant updated versions of Arm’s Mobile Studio profiling tools and Vulkan driver support, evidenced here by the Mali G2-Ultra NX MP16 configuration’s confirmed Vulkan 1.4 support. Whether major game engines like Unity and Unreal Engine ship day-one optimizations for neural graphics acceleration on this hardware is not yet confirmed and will likely depend on how quickly commercial devices actually ship with the new Mali GPU.
Developers working on on-device AI features, such as local language model inference or AI-assisted photo editing, are the more immediate audience for the CPU-side SME2 gains. A claimed 1.7x speedup on AI models with 38% less power is the kind of number that could shift where developers choose to run inference locally versus routing to a cloud API, though actual gains will vary heavily by workload and by how well a given app’s AI framework has been optimized for SME2.
Market Impact and Investor Reaction
Arm’s business model means this announcement is primarily a signal to its licensees rather than a direct product launch. The near-term financial impact, if any, would show up in future licensing and royalty revenue as chipmakers adopt CSS for Mobile 2 designs, a process that typically takes a year or more to reach commercial silicon. None of the sources reviewed for this article reported an immediate stock price reaction tied specifically to this announcement, and readers should be skeptical of any claim tying a specific share price move to the Mali G2-Ultra NX unveiling unless a named financial outlet reports it directly.
What is clear is the strategic stakes: Arm’s licensing business depends on convincing MediaTek, Samsung, and eventually Qualcomm’s non-custom lines to keep building on Arm reference designs rather than pushing further toward fully custom cores, as Qualcomm has done with Oryon and as Apple has done for over a decade with its in-house silicon. A well-received architecture like G2-Ultra NX, if it delivers on its neural graphics promise in real shipping phones, strengthens Arm’s case that its reference designs remain competitive rather than a budget fallback option. The pressure isn’t confined to phones either, since Nvidia’s push into edge AI hardware with products like the RTX Spark N1X shows how every chip vendor is racing to put more AI compute closer to the end user rather than routing it through the cloud.
Predictions: Where This Goes Next
Based on the confirmed roadmap and historical adoption patterns for Arm’s mobile compute subsystems, here’s what’s likely over the next 12 to 18 months.
- MediaTek will likely be the first vendor to ship a commercial chip using Mali G2-Ultra NX, given the Dimensity 9600 Pro pairing already spotted in Geekbench listings tied to the Vivo X500 Pro Max.
- Expect the first commercial phones built on CSS for Mobile 2 silicon to arrive in the first half of 2027, consistent with CNET’s framing of this as “2027 Android phones” technology rather than a 2026 product.
- Game engine vendors will likely announce Vulkan 1.4 and neural graphics API support updates before actual hardware ships broadly, since Arm typically pushes software tooling ahead of silicon availability.
- Qualcomm and Samsung’s next flagship SoC announcements will be scrutinized for whether they adopt any Arm C2 or Mali G2-Ultra NX blocks, or continue down fully custom CPU and GPU paths, a decision that will say a lot about how much pressure Arm’s new architecture is putting on competitors.
- Ray tracing benchmark comparisons between RTUv3-equipped Mali G2-Ultra NX phones and Qualcomm’s Adreno-based flagships will likely become a recurring talking point in 2027 flagship phone reviews once real hardware exists to test.
The Software Stack Behind the Silicon
Arm’s platform announcement bundled more than raw silicon design. The company described an updated developer tools stack that ships alongside CSS for Mobile 2, intended to give chip partners and app developers a common baseline for testing and optimizing against the new CPU and GPU IP before physical silicon is broadly available. This mirrors how Arm has handled prior CSS launches, where reference software and simulation tools precede actual chip tape-outs by months.
For developers evaluating whether to target SME2-specific code paths or wait for broader hardware availability, the practical answer right now is to wait for confirmed commercial silicon. Early Geekbench listings and pre-announcement GPU database entries, like the ones Notebookcheck has tracked for Mali G2-Ultra NX, are useful signals but are not the same as a shipping, reviewable device.
Frequently Asked Questions
What is Arm’s C2-Ultra CPU?
C2-Ultra is Arm’s new performance-focused mobile CPU core, part of the second-generation Arm C2 CPU cluster announced September 8, 2026. Arm describes it as its most powerful mobile CPU core, paired with the efficiency-focused C2-Pro core and two SME2 units for AI acceleration.
What is the Mali G2-Ultra NX GPU?
Mali G2-Ultra NX is Arm’s first AI-native mobile GPU, integrating dedicated neural network accelerators directly into its shader cores rather than as a separate chip block. Notebookcheck’s September 2026 GPU database entry shows that integration isn’t spread evenly across the die: in the MP16 configuration, only 8 of the 16 compute units are built with dedicated NX tensor cores. It also introduces third-generation ray tracing units (RTUv3) and a new execution engine.
Which phones will use Arm C2-Ultra and Mali G2-Ultra NX?
The clearest confirmed shipping device is the Xiaomi 18 Fold, launched September 8, 2026 with Mali G2-Ultra NX built into Xiaomi’s own Xring O3 chip. A second pairing has also surfaced in a Geekbench listing for the Vivo X500 Pro Max, which showed a 12-core Mali G2-Ultra NX GPU (labeled Mali-G2-Ultra-NX MC12) paired with a MediaTek Dimensity 9600 Pro chip. No other specific devices have been officially confirmed as of this writing.
When will phones with this new Arm platform launch?
Reports from CNET and other outlets indicate this platform targets 2027 Android phones, since chip designers typically need a year or more after an Arm IP announcement to tape out and ship commercial silicon.
How much faster is C2-Ultra than the previous generation?
Arm reports up to 15% higher single-thread performance and up to 12% higher multi-thread performance compared with the previous C1 generation, plus up to 1.7x faster AI model performance using two SME2 units while consuming 38% less power.
Is Mali G2-Ultra NX better than Qualcomm’s Adreno or Apple’s GPU?
There is no published head-to-head benchmark comparing Mali G2-Ultra NX against Qualcomm’s Adreno graphics or Apple’s in-house GPU designs. The three companies take different architectural approaches to AI acceleration, with Arm integrating neural accelerators into the GPU shader cores while Qualcomm and Apple keep dedicated NPU blocks separate.
What does CSS for Mobile 2 mean?
CSS stands for Compute Subsystem, Arm’s term for a pre-integrated bundle of CPU, GPU, and system interconnect IP that chip designers license as a starting point. CSS for Mobile 2 is the second generation of this bundle, combining the C2 CPU cluster, Mali G2-Ultra NX GPU, and an updated SI L2 interconnect.
Does Mali G2-Ultra NX support ray tracing?
Yes. It introduces third-generation ray tracing units, referred to as RTUv3, superseding the RTUv2 hardware used in the prior Mali generation, alongside a new execution engine for handling AI-native graphics workloads.
