Arm Unveils CSS for Mobile 2, Repositioning Mobile CPUs for On-Device Agent Orchestration

Translation Summary

Okay, here's my translation of the Arm CSS for Mobile 2 article. Based on my expertise, I've aimed for accuracy, preserving the original intent and technical nuance, and maintaining that journalistic flow:

On September 8th, Arm held its annual flagship event, Arm Everywhere China, in Shanghai, where it unveiled its second-generation mobile compute subsystem: CSS for Mobile 2. Designed for agentic AI and AI-native graphics, this computing platform integrates the Arm C2 CPU cluster, the Mali G2-Ultra NX GPU, and the SI L2 system interconnect, while combining system IP, software, and developer tools into a comprehensive platform.

Judging strictly by its name, CSS for Mobile 2 might seem like a routine iterative update. However, after listening to the entire presentation and discussing the technical details with Arm executives during interviews, we gained a fresh perspective on this new platform. With every technical upgrade in CSS for Mobile 2, Arm is essentially addressing a more fundamental question: As AI evolves from a Q&A assistant into an autonomous agent capable of understanding intent, planning tasks, and executing actions on behalf of users, what should a smartphone's computing platform look like?

Arm’s assessment is that mobile device performance hinges on three core dimensions: the execution speed of each task stage, the data transfer efficiency between different computing engines, and the system's coordinated orchestration across the entire workflow.

Accordingly, CSS for Mobile 2 coordinates optimizations at the system level and extends them to the physical implementation phase, allowing partners to evaluate power, performance, and area (PPA) alongside architectural design prior to SoC integration. It retains component flexibility—it can be adopted as a complete turnkey solution or combined with proprietary or third-party IP.

This reflects the consistent philosophy behind Arm CSS: packaging underlying computing technologies into an integrated, verified, and flexibly configurable foundation. In other words, partners no longer need to pour massive amounts of time and engineering resources into integrating foundational technologies. Instead, building on top of CSS, they can channel their resources into areas that deliver true differentiation—such as proprietary accelerators, memory architectures, interconnects, system technologies, and software capabilities—tailoring custom silicon to their target markets and workloads. By avoiding foundational pitfalls, partners benefit from faster development and reduced complexity, accelerating time-to-market for differentiated innovations while fully capitalizing on the strengths of the Arm ecosystem.

Let’s take a closer look at the details.

C2 CPU Cluster: Dual SME2 and Orchestration Engine

Today, AI agents do far more than run inference; they must maintain context, execute applications, coordinate diverse models and services, and autonomously carry out tasks with user authorization—all within the thermal and power constraints of a smartphone. The computing workload on a phone has shifted from "answering a question" to "autonomously executing an end-to-end workflow."

In this workflow, Arm believes the CPU has been redefined as the system's orchestration engine, responsible for maintaining context, scheduling workloads, and coordinating tasks.

The new C2 CPU cluster consists of Arm’s most powerful mobile CPU to date, the C2-Ultra, the efficiency-oriented C2-Pro, and dual SME2 engines. When running the latest AI models, the new CPU cluster delivers up to a 1.7x performance boost, a 15% increase in single-threaded performance, and up to a 38% reduction in power consumption at equivalent performance.

Looking closer, when running the Moonshine and Parakeet speech-to-text models, the C2-Ultra reduces latency by 40% compared to the previous-generation C1-Ultra, directly cutting down the wait time before an agent "responds." Memory retrieval performance during the memory phase improves by 41%, and the average instruction generation speed of small language models (SLMs) during inference increases by 25%. Taken end-to-end, overall workflow performance improves by 24%.

To clarify, the 15% improvement stems from general benchmarks like Geekbench 6.3, while the 1.7x gain corresponds to matrix computation throughput on specific AI models. The key driver here is SME2—this generation doubles the SME2 engines from one to two, doubling matrix compute capacity to achieve this leap. Additionally, this platform generation supports cutting-edge process nodes, including 2nm, allowing partners to deploy even more SME2 units on advanced processes to unlock further performance headroom for the future.

At this point, some might wonder: after years of smartphone makers engaging in an arms race over NPU TOPS, has the battleground shifted back to the CPU and memory subsystems in the agentic AI era?

This very question was raised by media during the interviews. In response, Daniel Lu, Director of Product Management for Edge AI CPU at Arm, noted that over the past decade, the industry's focus on different workloads has continually evolved alongside computing demands, constantly

On September 8, Arm hosted its annual flagship event, Arm Everywhere China, in Shanghai, where it unveiled its second-generation mobile compute subsystem: CSS for Mobile 2. Designed for agentic AI and AI-native graphics, this computing platform integrates the Arm C2 CPU cluster, the Mali G2-Ultra NX GPU, and the SI L2 system interconnect, while bringing together system IP, software, and developer tools into a comprehensive platform.

Judging strictly by its name, CSS for Mobile 2 might seem like a routine iterative update. However, after listening to the entire presentation and discussing the technical details with Arm executives in interviews, we came away with a fresh understanding of this new platform. With every architectural refinement in CSS for Mobile 2, Arm is essentially answering a more fundamental question: As AI evolves from a basic Q&A assistant into an autonomous agent capable of understanding intent, planning tasks, and executing actions on behalf of users, what should a smartphone's computing platform look like?

Arm’s assessment is that mobile device performance hinges on three core dimensions: the execution speed of each task stage, the data transfer efficiency between different computing engines, and the system's coordinated orchestration across the entire workflow.

Accordingly, CSS for Mobile 2 coordinates optimizations at the system level and extends them all the way to the physical implementation phase, allowing partners to evaluate power, performance, and area (PPA) alongside architectural design prior to SoC integration. It retains architectural flexibility—it can be implemented as a complete turnkey solution or integrated with proprietary or third-party IP.

This embodies the enduring philosophy behind Arm CSS: packaging foundational computing technologies into an integrated, verified, and flexibly configurable platform. In other words, partners no longer need to exhaust substantial time and engineering resources integrating baseline technologies. Instead, building on top of CSS, they can channel their resources where true differentiation is made—proprietary accelerators, memory architectures, interconnect technologies, system innovations, and software capabilities—tailoring custom silicon to their target markets and workloads. By avoiding foundational engineering hurdles, partners benefit from accelerated development schedules and reduced complexity, bringing differentiated innovations to market faster while fully leveraging the strengths of the Arm ecosystem.

Let’s take a closer look at the details.

C2 CPU Cluster: Dual SME2 and Orchestration Engine

Today, AI agents do far more than run inference; they must maintain context, execute applications, coordinate diverse models and services, and autonomously carry out tasks under user authorization—all within the thermal and power envelopes of a smartphone. The computing workload on a phone has shifted from "answering a question" to "autonomously executing an end-to-end workflow."

Arm believes that across this workflow, the CPU has been redefined as the system's orchestration engine, responsible for maintaining context, scheduling workloads, and coordinating tasks.

The new C2 CPU cluster comprises Arm’s most powerful mobile CPU to date, the C2-Ultra, the efficiency-focused C2-Pro, and dual SME2 engines. When running the latest AI models, this new CPU cluster delivers up to a 1.7x performance improvement, a 15% increase in single-threaded performance, and up to a 38% reduction in power consumption at equivalent performance.

Delving deeper, when running the Moonshine and Parakeet speech-to-text models, the C2-Ultra achieves a 40% latency reduction compared to the previous-generation C1-Ultra, directly minimizing the delay before an agent responds verbally. Memory retrieval performance during the memory phase increases by 41%, and the average instruction generation speed of small language models (SLMs) during inference improves by 25%. On an end-to-end basis, overall workflow performance rises by 24%.

To clarify, the 15% improvement stems from general-purpose benchmarks such as Geekbench 6.3, whereas the 1.7x gain reflects matrix computation throughput on specific AI models. The credit goes to SME2: doubling the SME2 engines from one to two in this generation doubles matrix compute capacity, enabling this leap. Furthermore, this platform generation supports advanced process nodes down to 2nm, allowing partners to deploy additional SME2 units on cutting-edge nodes to secure further performance headroom for the future.

At this point, one might wonder: after years of smartphone makers engaging in an arms race over NPU TOPS, has the center of competition shifted back to the CPU and memory subsystem in the era of agentic AI?

This question was indeed raised during media interviews. Daniel Lu, Director of Product Management for Edge AI CPU at Arm, explained that over the past decade, the industry's focus on different workloads has continually evolved alongside computational demands, always seeking the optimal balance of compute power for varying devices and use cases. Current flagship NPUs deliver roughly 100 to 200 TOPS, while an SME-equipped CPU cluster provides an equivalent of about 5 to 6 TOPS. Workloads with extreme compute density are better suited for GPUs or NPUs, whereas the CPU’s battleground lies with small language models that comfortably fit within local compute budgets.

He particularly emphasized that CPUs prioritize latency over peak throughput, noting that blindly chasing peak performance can ultimately compromise latency.

In addition, to overcome the prevalent memory-bound bottlenecks on edge devices, the C2 CPU cluster adopts a differentiated cache topology featuring private L2 caches paired with a large shared L3 cache, leveraging this two-tier cache capacity to reduce DRAM access frequency. Lu acknowledged that memory-bound constraints on mobile devices are often more acute than compute-bound ones, and the cache topology represents their primary architectural solution to this challenge.

Another critical yet frequently overlooked component is the orchestration layer. It manages task states, consolidates context, and determines which computing resource executes each phase of the workflow—whether that is a local application, an on-device model, or a cloud service. The CPU centrally coordinates permission management and task execution throughout. In the age of AI agents, the role of the CPU has evolved from simply computing fast to orchestrating effectively.

Regarding core configurations, Lu revealed that the C2 CPU cluster supports up to 14 cores, giving partners the flexibility to configure silicon according to their specific workloads. For example, some partners evaluated agentic workloads, concluded that at least ten cores were required, and chose a 10-core layout. Arm does not dictate these decisions; it delivers the architectural flexibility to its partners.

Ecosystem adoption of SME2 is also gathering pace. James McNiven, Vice President of Client Computing for Edge AI at Arm, noted that virtually all flagship smartphones—across both Android and iOS—have now adopted SME2 technology. The ecosystem surrounding SME2 includes Alipay, Google AI Edge Gallery, OPPO, and vivo, among others. At the event, Alibaba’s Tongyi Qianwen team demonstrated how the Qwen-Audio family runs efficiently on-device via Arm CPUs and SME2, delivering the required performance and precision for high-fidelity speech generation.

Mali G2-Ultra NX: AI-Native Graphics Architecture, Embedding Neural Networks into Shaders

Moving from the CPU to the GPU: the Mali G2-Ultra NX is Arm’s first AI-native Mali GPU. Its defining architectural shift is the direct integration of dedicated neural network accelerators into the GPU shader cores. This allows neural graphics workloads to run concurrently with graphics and compute tasks while sharing the GPU's memory subsystem, coherent caches, and control structures, minimizing data movement between separate compute engines.

During interviews, Deyan Lazarov, Senior Product Manager for Edge AI at Arm, clarified an important technical detail: the system employs Convolutional Neural Networks (CNNs), not Transformers. Rather than hallucinating imagery from scratch like generative models, the system takes the native frames rendered by the GPU and utilizes CNNs for super-resolution and detail reconstruction. For titles with unique visual styles, developers can also retrain the models on their own proprietary game assets.

The Mali G2-Ultra NX supports three neural acceleration technologies: Neural Super Sampling (NSS), which reconstructs high-resolution images from lower-resolution inputs; Neural Frame Rate Upscaling (NFRU), which boosts frame rates by generating interpolated frames; and Neural Super Sampling and Denoising (NSSD), which unites super-resolution and denoising for complex ray-traced scenes.

In Light and Shadow: Rebirth, a tech demo developed on Unreal Engine in collaboration with Sumo Digital, the combined NFRU and NSSD solution delivered up to a 4x improvement in performance efficiency compared to conventional rendering, cut external memory traffic by up to 70%, and enabled desktop-class features like Unreal Engine’s MegaLights to run on mobile devices.

Addressing latency concerns, Lazarov broke down the timing budget: targeting 30 FPS allows a per-frame window of roughly 33 milliseconds. Developers must complete both the native frame render and the AI frame generation within this window, outputting a visual experience equivalent to 60 FPS while employing frame pacing mechanisms to absorb rendering fluctuations. While this generation of NFRU supports inserting one AI-reconstructed frame between two rendered frames, Arm's roadmap already encompasses multi-frame generation capabilities.

To grant developers greater latitude in balancing trade-offs, Arm also introduced multiple model quality tiers, allowing fine-tuning between visual fidelity and performance. Lazarov explained that Light and Shadow: Rebirth natively renders at a lower resolution and applies upscaling and reconstruction during the frame generation phase; when rendering demand surges, native resolution can be dialed back even further.

Traditional graphics performance has received equal attention. The all-new execution engine represents the most substantial instruction set architecture (ISA) upgrade across the last seven generations of Arm Mali GPUs. It provides up to 2x the registers per warp compared to the prior generation and utilizes dynamic register allocation to minimize register spills. Furthermore, the third-generation ray tracing unit introduces hardware support for Opacity Micromaps (OMM), which can elevate frame rates by 30% and reduce ray tracing workloads by up to 70%.

Compared to the previous generation, the GPU delivers up to a 24% uplift in benchmark performance, up to a 14% improvement in non-AI gaming performance, up to a 13% reduction in DRAM traffic under ray-traced workloads, and sustained performance at up to 120 FPS.

During the interviews, media also drew comparisons to desktop frame generation techniques. McNiven acknowledged that smartphones fundamentally differ from PCs: hemmed in by stringent power and thermal limits with finite resources, they require specialized architectural design and dedicated optimization engineered specifically for mobile environments.

Another key differentiator for Arm is its open philosophy: the models are open, allowing smartphone manufacturers to validate, optimize, and customize them according to their individual product roadmaps. Arm also recently announced a collaboration with Tencent Games to explore next-generation technologies, including Neural Dynamic Global Illumination.

Finally, the SI L2 system interconnect serves as the connective backbone of the platform. Supporting multiple memory standards—including LPDDR6, LPDDR5X, and LPDDR5—it delivers substantial architectural enhancements in cache coherency and large-model optimizations over its predecessor.

McNiven noted that Arm’s role is to ensure that SI L2 seamlessly accommodates emerging memory standards as they enter the market. Even when partners choose not to adopt SI L2, they typically implement comparable system-level optimizations to reduce latency and support next-generation memory architectures.

Beyond Silicon: A Fabric Bringing Intelligence Everywhere

From a hardware standpoint alone, the CSS for Mobile 2 platform presents an exceptionally complete picture. Yet industry observers might look further: beyond the technology itself, CSS