Arm has unveiled C2-Ultra, its new top-performance mobile CPU, integrated into the C2 CPU Cluster within the CSS for Mobile 2 platform. The company says the new core improves single-thread performance by up to 15% over the previous generation and speeds up certain AI workloads by as much as 1.7x. The more interesting shift, however, is in the approach: Arm is positioning the CPU to coordinate AI agents that combine local models, applications, memory, GPU, accelerators, and cloud services.
Arm C2-Ultra in 20 seconds
- C2-Ultra replaces C1-Ultra as the top-performance core of Arm’s new mobile platform.
- Arm measures up to 15% more single-thread performance and 12% faster app launches.
- The C2 CPU Cluster packs up to two SME2 units to accelerate on-device AI.
- The company cites up to 1.7x more performance on select AI models and a 24% shorter time for a full agentic workflow.
C2-Ultra doesn’t arrive alone. It’s part of the C2 CPU Cluster, which pairs C2-Ultra cores for the most latency-sensitive tasks with C2-Pro cores for sustained workloads and two Scalable Matrix Extension 2 (SME2) units. That cluster, in turn, is one of the building blocks of CSS for Mobile 2, where Arm also adds the new Mali G2-Ultra NX GPU and the SI L2 system interconnect.
The architecture shows where Arm wants to steer the conversation around mobile AI. A neural processing unit (NPU) remains well suited to running certain models, but an agent that works through multiple steps needs considerably more than inference. It has to prepare context, run code, launch applications, retrieve information, manage permissions, and decide where each part of the task should be processed.
For Arm, a significant share of that work falls back on the CPU.
C2-Ultra Improves Single-Thread Performance by 15% and Efficiency by Up to 38%
Arm presents C2-Ultra as its most powerful mobile CPU to date. According to the company’s internal testing, carried out between March and August 2026, it delivers up to 15% more single-thread performance than C1-Ultra. Real-world results will depend on the final system-on-chip (SoC) design, clock speeds, memory, and thermal limits of each device.
The company also shares other figures for the C2 CPU Cluster:
| Metric announced by Arm | Improvement over the previous generation |
|---|---|
| Single-thread performance | Up to +15% |
| Web browsing | +15% |
| App launch | +12% |
| Cluster multi-thread performance | +12% |
| Performance on select AI models | Up to 1.7x |
| Full reference agentic workflow | +24% |
| AI performance with two SME2 units vs. C1 CPU Cluster | Up to +70% |
These are Arm’s own figures, not independent benchmarks. The improvement percentage also varies depending on the workload measured, so the announced 1.7x for AI shouldn’t be read as a blanket speedup for every model running on C2-Ultra.
There’s another notable figure that gets less prominence in the initial presentation: Arm claims that C2-Ultra can use up to 38% less power to deliver the same level of performance as C1-Ultra.
On a smartphone, that performance-to-power ratio matters as much as peak performance. A processor can complete an operation quickly, but persistent AI workloads introduce a different problem: keeping models, retrieval, and applications running for longer without exceeding the device’s thermal limits.
That’s where C2-Pro comes in.
Arm reserves C2-Ultra for immediate response and foreground tasks, while C2-Pro is designed to sustain longer workloads more efficiently. C2-Pro uses the same efficiency-oriented microarchitecture as C1-Pro, though built on newer manufacturing process technology.
SME2 Turns the CPU Into Another Accelerator for AI Models
One of the most important technical pieces of the C2 CPU Cluster is Scalable Matrix Extension 2.
SME2 extends the Arm architecture with instructions and structures built for matrix operations. These operations show up constantly in machine learning, so they let certain AI workloads run far more efficiently than relying solely on the CPU’s general-purpose instructions.
CSS for Mobile 2 can integrate two SME2 units, double the capacity of the previous configuration.
Arm says this combination delivers up to 70% more performance than the C1 CPU Cluster on certain AI tasks. Comparing C2-Ultra with C1-Ultra on select models, the maximum announced improvement reaches 1.7x.
The goal isn’t to turn the CPU into a replacement for an NPU or GPU.
The idea is to have more places to run each part of an AI application.
A small model dedicated to speech recognition, classification, retrieval, or planning can be light enough to run efficiently through SME2. A heavier model can go to the NPU. Neural graphics tasks can run on the Mali G2-Ultra NX, and a particularly complex request may still need a model hosted in the cloud.
The CPU coordinates that journey.
This heterogeneous division of labor explains Arm’s bet on so-called agentic AI better than simply comparing TOPS figures between processors.
A Mobile Agent Needs Much More Than Running an LLM
Today’s assistants generally follow a fairly simple pattern: they receive an input, run one or more models, and return a result.
An agent can sustain a much longer sequence.
For example, a request to find a restaurant, check the calendar, and set up a reservation could require speech recognition, retrieval of preferences, reasoning, querying apps, web browsing, and executing an action.
Not everything in that journey is inference.
Arm describes the CPU as the component responsible for keeping track of the task’s state, preparing context, and coordinating where each stage runs. Some operations can stay entirely on the device, while others rely on external services.
To measure this, the company built a reference agentic workflow that combines voice processing, memory retrieval, reasoning, app execution, and web browsing.
In that test, C2-Ultra paired with two SME2 units completes the process 24% faster than the previous generation, according to Arm.
That’s a more useful metric than measuring an isolated inference alone, though it’s still a test designed by the manufacturer, and it will take commercial devices to see how it plays out in real applications.
CSS for Mobile 2 Brings Together CPU, GPU, and Interconnect
C2-Ultra is also just one part of a larger announcement.
The new CSS for Mobile 2 is a Compute Subsystem (CSS) platform that bundles components manufacturers can use to build their own mobile SoCs.
Its elements include the C2 CPU Cluster, the new Mali G2-Ultra NX GPU, the SI L2 interconnect system, physical implementations, and software tools.
Mali G2-Ultra NX brings neural accelerators directly into the graphics architecture. Arm is looking to apply machine learning to rendering and graphics techniques this way, while the C2 CPU Cluster handles general-purpose workloads and part of the on-device AI.
SI L2 rounds out the architecture by managing data movement and coherence across the different compute resources.
This approach also gives manufacturers some flexibility. The components can be integrated as part of CSS for Mobile 2 or used within designs that incorporate their own IP or third-party IP.
That’s why the C2-Ultra announcement isn’t the same as launching a commercial processor that can be bought directly. Arm provides IP and designs that its partners then integrate into their chips.
Clock speeds, core count, memory, manufacturing process, and final configuration will depend on those SoCs.
Software Will Matter as Much as SME2
Adding AI instructions to a CPU is of little use if every developer has to program them by hand.
Arm is trying to avoid that problem through KleidiAI, framework integrations, and its new AI Portal. The goal is for applications to benefit from SME2 through libraries and already-adapted paths rather than requiring every project to write its own custom kernels.
The company says SME2 already appears in Android and iOS devices and cites an ecosystem that includes Alipay, Google AI Edge Gallery, OPPO, and vivo. That gives it a software base on which to extend C2’s capabilities.
The shift also matters for mobile developers. Until now, much of the conversation around on-device AI has revolved around how many TOPS an SoC’s NPU offers.
C2-Ultra proposes a somewhat different architecture: the CPU, GPU, and specialized accelerators can all take part in the same AI application, each handling the parts that best match its strengths.
The push mirrors a similar shift Arm is making on the data center side, where it is targeting 512-core CPUs built for the AI era.
It remains to be seen how C2-Ultra will perform in commercial silicon and which manufacturers will adopt the different CSS for Mobile 2 configurations. The figures published so far come from Arm and compare its new generation with the previous one under its own methodology.
But the design points to a clear technical shift. The smartphone’s CPU is no longer just there to open apps, run JavaScript, or keep Android running. Arm wants it to also coordinate a new generation of applications where multiple models, accelerators, and services can be working at the same time.
Under that framing, on-device AI stops being just a matter of running a fast model. The problem becomes how to distribute and coordinate all the work needed for that model to actually act on the device.
