Arm used its hardware newsroom on September 8, 2026 to introduce the Arm C2-Ultra CPU and the Mali G2-Ultra NX GPU, the two building blocks of what the company is calling its next mobile compute platform. The announcement lands at a moment when phone and tablet chip vendors are racing to bolt dedicated AI hardware onto every layer of the silicon stack, and Arm’s pitch this time is that the CPU and GPU should carry that load together instead of leaving it to a separate neural processing unit.
For engineers and hardware buyers, the practical question is simple: does this generation change what a flagship phone can actually do on-device, or is it another round of incremental IP licensing news that only matters to chip designers. The numbers Arm has published, plus the architectural choices behind them, suggest this is closer to a real inflection point for how mobile GPUs handle AI workloads, even if plenty of details about pricing, partner silicon, and shipping dates remain open.
What Arm Announced on September 8
The core of the announcement is two pieces of licensable IP. The Arm C2-Ultra CPU core adds support for SME2 (Scalable Matrix Extension 2), which Arm describes as bringing “a new level of responsive, efficient on-device AI to the CPU,” according to the company’s newsroom post. Paired with it is the Mali G2-Ultra NX GPU, which Arm calls its “first AI-native Mali GPU,” per the same announcement.
Arm’s product manager Deyan Lazarov was named in newsroom coverage of the launch on September 8, 2026, describing the Mali G2-Ultra NX as a rework of how the GPU pipeline handles neural workloads rather than a simple clock-speed bump. The framing matters because Arm ships IP, not chips. Arm publishes its CPU IP portfolio and GPU IP portfolio for licensees to review, but neither the C2-Ultra nor the Mali G2-Ultra NX will appear in a retail device on their own. Instead, licensees like Qualcomm, MediaTek, Samsung, and others build these designs into system-on-chip products that then ship inside phones, tablets, and other devices over the following 12 to 18 months.
That licensing model is why Arm news cycles tend to run well ahead of retail availability. Today’s announcement establishes the reference architecture. Actual silicon carrying these cores typically surfaces in flagship Android chipsets the following year, which means most of the real-world performance story will be told by benchmark leaks and OEM chip unveilings well into 2027.
Arm C2-Ultra: The CPU Side of the Story
On the CPU side, Arm is citing up to 15% higher single-thread performance for the C2-Ultra core, or alternatively up to 38% lower power draw at the same performance level as the prior generation, according to Arm’s own published figures. Those two numbers describe the same underlying silicon budget spent two different ways: a phone maker chasing benchmark scores can push clocks and grab the 15% uplift, while a maker prioritizing battery life can hold performance flat and pocket the power savings instead.
SME2 support is the more consequential change for developers. Scalable Matrix Extension instructions let the CPU execute matrix-multiply operations, the core math behind most neural network inference, without routing that work through a separate accelerator. For app developers running small language models, on-device transcription, or local image processing, that means the CPU itself becomes a viable execution target for AI workloads that previously needed a dedicated NPU or a round trip to the GPU.
This is not a wholesale redesign of the Arm instruction set. It is closer to the SVE2/SME progression Arm has been building toward for several generations, extended further into the mobile CPU line. What is new is that Arm is positioning it explicitly around what the company terms agentic AI workloads, where a phone may need to run several AI-driven processes concurrently rather than a single chatbot query at a time.
Mali G2-Ultra NX: An AI-Native GPU Pipeline
The GPU changes are where Arm is making its bigger claim. The Mali G2-Ultra NX introduces dedicated neural accelerators built directly into the shader cores and graphics pipeline, rather than treating AI inference as a workload bolted onto the side of the GPU. Arm says this design can deliver up to 4x higher performance per watt on neural graphics workloads compared with prior Mali generations.
On traditional graphics work, unrelated to AI, Arm cites a more modest 14% performance increase. The gap between those two figures, 4x for neural graphics versus 14% for conventional rendering, is the clearest signal of where Arm actually redesigned the architecture this generation. The company built new dedicated paths for AI inference and left the traditional rasterization pipeline largely as an incremental refinement.
Ray tracing gets its own upgrade with a third-generation Ray Tracing Unit. Arm says the new RT hardware, combined with a new feature called Opacity Micromaps, can cut ray-tracing workload cost by up to 70% and lift frame rates by up to 30% in tested scenarios. Opacity Micromaps work by letting the GPU classify small regions of a triangle as fully opaque, fully transparent, or requiring a shader check, which trims the number of expensive any-hit shader invocations during ray traversal. That is the same general technique Nvidia introduced on desktop RTX hardware, now arriving in a mobile-class ray tracing unit.
A separate memory-bandwidth claim, detailed in technical coverage of the new IP, rounds out the graphics story. Arm says the new ray-tracing unit delivers up to 13% lower DRAM traffic on leading ray-tracing benchmarks. On a phone, where DRAM bandwidth is shared, power-constrained, and directly tied to thermal headroom, cutting memory traffic matters as much as raw compute gains. A GPU that hits a frame-rate target while moving less data through memory runs cooler and sustains that performance for longer before the SoC throttles.
Why “AI-Native” GPU Design Is the Real Shift
Mobile GPUs have supported general compute workloads for years through APIs like Vulkan compute and OpenCL, and NPUs have handled dedicated AI inference in parallel silicon blocks since roughly 2017. What Arm is describing with the Mali G2-Ultra NX is different: neural accelerators embedded inside the same shader cores that render graphics, sharing the GPU’s memory system, coherent cache, and control logic rather than sitting off to the side as a separate compute island.
That matters for a specific class of workload that is becoming more common: neural graphics, where AI models generate or upscale portions of a frame in real time, similar in spirit to how DLSS and FSR use machine learning to reconstruct image detail on desktop GPUs. When the neural inference step happens inside the same pipeline as the rasterization step, the GPU avoids shipping intermediate frame data out to a separate NPU and back, which is exactly the kind of round trip that burns power and adds latency on a battery-constrained device.
The tradeoff is architectural complexity. Building neural accelerators into the shader core pipeline is a bigger engineering lift than adding a standalone block, and it commits Arm to a specific bet about how mobile AI graphics workloads will evolve. If game engines and camera pipelines increasingly interleave AI inference with traditional rendering, as they already do on PC, that bet pays off. If most AI workloads stay CPU or NPU-bound instead, the GPU-integrated approach delivers less relative benefit.
Arm C2-Ultra and Mali G2-Ultra NX: Key Specs at a Glance
| Component | Claimed Improvement | Compared To | Source |
|---|---|---|---|
| Arm C2-Ultra CPU | Up to 15% higher single-thread performance | Prior generation Arm mobile CPU core | Arm newsroom, Sept. 8, 2026 |
| Arm C2-Ultra CPU | Up to 38% lower power at equal performance | Prior generation Arm mobile CPU core | Arm newsroom, Sept. 8, 2026 |
| Mali G2-Ultra NX GPU | Up to 4x performance-per-watt on neural graphics | Prior Mali GPU generation | Arm newsroom, Sept. 8, 2026 |
| Mali G2-Ultra NX GPU | 14% higher traditional graphics performance | Prior Mali GPU generation | Arm newsroom, Sept. 8, 2026 |
| Mali G2-Ultra NX Ray Tracing Unit (3rd-gen) | Up to 70% lower ray-tracing workload cost, up to 30% higher frame rate | Prior-gen Mali ray tracing unit | Arm newsroom, Sept. 8, 2026 |
| Mali G2-Ultra NX Ray Tracing Unit (3rd-gen) | Up to 13% lower DRAM traffic | Leading ray-tracing benchmarks, prior gen | Arm newsroom, Sept. 8, 2026 |
How This Fits Arm’s Release History
Arm has shipped a new flagship Mali GPU roughly once a year for the past several generations, each time layering in graphics features that first proved out on desktop hardware. Hardware ray tracing support arrived in Mali GPUs a few generations back, well after Nvidia and AMD normalized it on desktop cards, and the pattern with AI-native shader design looks similar: mobile silicon adopting a technique after it is already established on PC GPUs, then adapting it to a tighter power envelope.
What is new this cycle is the framing. Previous Mali generations pitched incremental graphics or efficiency wins as the headline. This announcement leads with AI-native design and agentic AI workloads as the organizing principle for both the CPU and GPU changes, which tracks with how the entire chip industry has recast its roadmaps around on-device AI inference over the last two years. The gaming and creative-workload benefits, like the ray tracing gains, are presented as a secondary payoff of an AI-first redesign rather than the primary goal.
Competitive Landscape: Where Arm’s IP Fits
Arm does not compete directly with chipmakers since it licenses core designs rather than selling finished silicon, but the C2-Ultra and Mali G2-Ultra NX still enter a crowded field of approaches to on-device AI. Qualcomm’s Snapdragon line pairs custom Arm-based CPU cores (in recent generations, Qualcomm’s own Oryon design rather than stock Arm cores) with a dedicated Hexagon NPU. Apple’s Silicon lineup uses custom CPU and GPU cores with a separate Neural Engine block. Samsung’s Exynos chips mix Arm reference CPU cores with AMD RDNA-derived GPU designs in some generations.
The distinction Arm is drawing with the Mali G2-Ultra NX is that neural acceleration lives inside the GPU shader cores themselves, rather than in a physically and logically separate NPU block, which is the architecture Apple, Qualcomm, and most Exynos designs currently use. Whether that integrated approach beats a well-tuned discrete NPU on real workloads will depend heavily on how software frameworks like TensorFlow Lite, ONNX Runtime, and each OEM’s own AI stack choose to schedule work across the CPU, GPU, and NPU on a given chip. Arm licensees are not obligated to use the reference Mali GPU at all. MediaTek and Samsung both maintain the option to license Mali IP or build custom graphics silicon, and Qualcomm has used its own Adreno GPU architecture for years rather than Mali, even as it recently raised chip prices amid AI-driven demand.
The broader chip industry backdrop makes Arm’s timing notable. AMD’s Ryzen AI Max PRO 400 pushed AI-capable memory capacity on the PC side earlier this year, Nvidia’s RTX Spark platform has drawn OEMs into compact AI-focused desktop hardware, and Intel’s Nova Lake roadmap points to a similar AI-first framing on its next Core Ultra generation. Even process-node announcements like Fujitsu’s 2nm Monaka CPU are being pitched around AI workload density rather than clock speed alone. Arm’s mobile-focused announcement fits the same industry-wide pattern of chipmakers reframing every product tier around on-device AI capability.
| Chip Platform | CPU Approach | AI Acceleration Location | GPU |
|---|---|---|---|
| Arm CSS with C2-Ultra / Mali G2-Ultra NX | Arm reference cores with SME2 | CPU matrix extensions + GPU-integrated neural accelerators | Mali G2-Ultra NX (reference) |
| Qualcomm Snapdragon (recent gens) | Custom Oryon cores | Dedicated Hexagon NPU block | Adreno (custom) |
| Apple Silicon | Custom Apple CPU cores | Dedicated Neural Engine block | Custom Apple GPU |
| Samsung Exynos (select gens) | Arm reference or custom cores | Dedicated NPU block | Arm Mali or AMD RDNA-derived (varies by generation) |
What This Means for Developers
For mobile developers, the practical near-term impact is limited: the C2-Ultra and Mali G2-Ultra NX are IP announcements, not shipping chips, so there is nothing to target in production yet. The longer-term implication is a widening set of places an app’s AI inference workload can actually run once C2-Ultra and Mali G2-Ultra NX-based silicon reaches devices. An app that currently forces all inference onto a dedicated NPU block, or falls back to cloud inference when the NPU is busy, will eventually have the option to spill matrix-heavy work onto the CPU through SME2 or onto the GPU’s neural accelerators, depending on which resource has headroom at that moment.
That flexibility mostly benefits frameworks that already abstract hardware placement, like ONNX Runtime’s execution providers or Google’s LiteRT, since those toolchains can route a given operation to whichever backend the underlying SoC exposes without app-level code changes. Developers hand-tuning inference paths for a specific NPU vendor’s SDK will need to wait for OEM chips built on this platform before there is anything concrete to test against.
Gaming and Graphics Implications
Mobile gaming is the other obvious beneficiary, particularly the ray tracing gains. A 70% reduction in ray-tracing workload cost, if it holds up in shipping silicon, is large enough to change which games can realistically offer ray-traced lighting or reflections on a phone at all, rather than as a novelty toggle that tanks frame rate and battery life. Combined with the claimed 30% frame-rate increase, engines like Unreal Engine and Unity could extend mobile ray tracing support beyond the small number of showcase titles that currently bother with it.
The neural graphics angle also opens the door to mobile equivalents of desktop upscaling technology. If a phone’s GPU can run AI-based frame reconstruction natively inside the shader pipeline rather than as a bolt-on pass, that’s a more efficient path to something resembling DLSS or FSR on a handset, letting games render at a lower internal resolution and upscale with less power overhead than running the same reconstruction model through a separate compute pass.
What Remains Unconfirmed
Several details that typically follow an Arm IP announcement have not surfaced yet. Arm has not published pricing for licensing the C2-Ultra or Mali G2-Ultra NX, which is normal since Arm’s licensing terms with chipmakers are generally confidential and never public regardless. No specific shipping dates or partner chip names have been confirmed for silicon built on this platform. Historically, Arm’s mobile CPU and GPU announcements precede shipping SoCs by roughly a year, with flagship phones carrying the resulting chips arriving in the following calendar year’s device cycle.
Reporting on chip leaks ahead of today’s announcement, including Geekbench listings referencing Mali G2-Ultra NX configurations, suggests some chipmakers have early silicon in testing, but Arm’s own announcement does not confirm any specific partner products, launch windows, or device names. Readers should treat any specific phone model or release date tied to this platform as speculation until an OEM confirms it directly.
Predictions: Where This Goes Next
- Expect the first C2-Ultra and Mali G2-Ultra NX-based SoCs from major licensees to surface in benchmark leaks well before official unveilings, following the pattern set by prior Mali generations.
- Flagship Android phones carrying this platform are more likely to arrive in the 2027 device cycle than in 2026, consistent with Arm’s typical IP-to-silicon lead time.
- Game engine vendors will likely highlight mobile ray tracing more aggressively in marketing once shipping chips demonstrate the claimed 70% workload reduction in independent benchmarks, rather than Arm’s own figures.
- Competing NPU-centric designs from Qualcomm and Apple will face pressure to publish their own performance-per-watt comparisons for neural graphics workloads as GPU-integrated AI acceleration becomes a marketing differentiator.
- Expect scrutiny of Arm’s efficiency claims once independent reviewers get hands-on silicon, since prior generations’ headline percentages have sometimes proven optimistic once measured outside Arm’s own reference test conditions.
The Bigger Picture for Arm’s Business
This announcement continues Arm’s push to be seen as an AI infrastructure company rather than strictly a mobile CPU licensor, a repositioning that has accelerated since the company’s 2023 IPO. Framing the C2-Ultra and Mali G2-Ultra NX around agentic AI workloads, rather than simple year-over-year performance gains, mirrors language Arm has used in its data center and PC roadmap announcements over the past year. The company is betting that AI capability, not raw clock speed, is now the primary axis licensees and end users judge a new chip generation on.
That bet carries real business stakes. Arm’s revenue model depends on royalties tied to the volume and value of chips built on its architecture, so a generation that chipmakers view as a must-have AI upgrade, rather than an optional refresh, has a direct path to higher per-chip royalty rates. Whether the C2-Ultra and Mali G2-Ultra NX clear that bar will become clear only once shipping devices let independent reviewers test the AI and ray-tracing claims against real workloads, rather than Arm’s own reference benchmarks.
Frequently Asked Questions
What is the Arm C2-Ultra?
The Arm C2-Ultra is a new mobile CPU core Arm announced on September 8, 2026, adding SME2 support for on-device AI matrix computation. Arm says it delivers up to 15% higher single-thread performance or up to 38% lower power at equal performance versus the prior generation.
What is the Mali G2-Ultra NX?
The Mali G2-Ultra NX is Arm’s new mobile GPU, described by the company as its first AI-native Mali GPU. It integrates dedicated neural accelerators directly into its shader cores and adds a third-generation Ray Tracing Unit with Opacity Micromap support.
When will phones with the Arm C2-Ultra and Mali G2-Ultra NX ship?
Arm has not confirmed specific shipping dates or partner devices. Based on Arm’s typical IP-to-silicon timeline, chips using this platform are more likely to appear in flagship phones during the 2027 device cycle than before the end of 2026.
How much faster is the Mali G2-Ultra NX at ray tracing?
Arm says its third-generation Ray Tracing Unit can cut ray-tracing workload cost by up to 70% and increase frame rates by up to 30% in tested workloads, alongside up to 13% lower DRAM traffic on leading ray-tracing benchmarks. These are Arm’s own reference figures and have not yet been independently verified on shipping silicon.
Is the Mali G2-Ultra NX the same as an NPU?
No. Rather than adding a separate neural processing unit block, Arm built dedicated neural accelerators directly into the Mali G2-Ultra NX’s shader cores and graphics pipeline, letting neural graphics workloads share the GPU’s existing memory system and cache instead of running on isolated compute hardware.
Which companies will use the Arm C2-Ultra and Mali G2-Ultra NX?
Arm licenses this IP to chipmakers rather than selling finished chips itself. No specific partner companies or products have been confirmed for this platform as of the September 8, 2026 announcement, though Arm’s mobile CPU and GPU designs are typically adopted by licensees including MediaTek, Samsung, and other Android chipmakers in subsequent SoC generations.
Does this replace Qualcomm or Apple’s own chip designs?
No. Qualcomm uses its own custom Oryon CPU cores and Adreno GPU rather than Arm’s reference designs, and Apple builds fully custom silicon. The C2-Ultra and Mali G2-Ultra NX are reference IP primarily aimed at licensees like MediaTek, Samsung, and other Android chipmakers who build directly on Arm’s designs.
What does SME2 do on the Arm C2-Ultra?
SME2, or Scalable Matrix Extension 2, adds CPU instructions optimized for matrix-multiply operations, the core computation behind most AI model inference. It lets the CPU handle certain AI workloads directly rather than routing them exclusively to a GPU or NPU.
