Agentic AI changes the mobile experience from answering questions to turning intent into action through gathering relevant context, planning actions, and coordinating applications and services on a user’s behalf.
These agentic workflows, alongside increasingly sophisticated AI-native graphics, place new demands on mobile compute. Mobile performance now depends on how quickly each stage can execute, how efficiently data moves between compute engines, and how effectively the system coordinates activities across the complete task.
Meeting these demands requires a compute platform designed and optimized as a complete system. Our new Arm CSS for Mobile 2 platform is designed for that shift, delivering an AI-native compute foundation powered by two new best-in-class compute engines built for the AI era. The Arm C2 Ultra CPU with Arm Scalable Matrix Extension 2 (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.
A platform designed around the complete agentic AI workload #
As agentic AI engages several parts of a mobile system within the same interaction, processing, model execution, memory access and application control must remain coordinated as the device interprets a goal and carries it through to completion. This system-level approach shapes how CSS for Mobile 2 is assembled and delivered. CPU, GPU, system IP, software and developer tools are brought together as a complete platform, allowing performance, efficiency and data movement to be considered across the entire workflow as a whole.
This heterogeneous approach supports workloads that increasingly span multiple compute resources, with the CPU providing the programmable compute and orchestration that coordinates activity across the system. Sustained performance and power efficiency help keep those repeated and concurrent AI interactions responsive within mobile power and thermal constraints.
The integrated platform preserves flexibility. Individual components can be used independently or combined within a broader CSS configuration alongside custom and third-party IP, giving partners the flexibility to shape systems around their product, performance and market requirements.
CSS for Mobile 2 also extends system optimization into physical implementation, allowing performance, power and area to be considered alongside the architecture before partners begin system-on-chip (SoC) integration.
Orchestrating agentic AI with the Arm C2 CPU cluster #
Within CSS for Mobile 2, the Arm C2 CPU cluster provides the programmable compute and orchestration needed to coordinate an agentic workflow from intent through to action. Combining C2-Ultra and Arm C2-Pro CPUs with two SME2 units, the cluster advances the everyday processing used across mobile applications while expanding the local AI capability available for speech, retrieval and model execution. This doubles SME2 capabilities compared with the previous-generation configuration and provides greater AI compute headroom for latency-sensitive on-device workloads.
C2 performance highlights include:
- Up to 1.7x performance improvement across the latest AI models;
- Up to 15 percent higher single-thread performance;
- 15 percent faster web browsing;
- 12 percent faster application launch; and
- 12 percent higher multi-thread performance at the cluster level, supporting parallel activity across retrieval, model preparation and application execution.
These gains matter in every stage of the agentic workflow. Higher single-thread performance reduces the latency of individual decisions and system events, while higher multi-thread performance helps searches, models and applications progress concurrently. Faster browsing and application launch shorten the time to access services and carry out actions.
At the center of this workflow is an orchestration layer that maintains task state, assembles relevant context and determines where each stage should execute. Depending on the workload, that can involve local applications, on-device models, other compute resources or cloud services, with the CPU coordinating permissions, decisions and the progression of the task.
Across a representative end-to-end agentic workflow spanning speech processing, memory retrieval, reasoning, application execution and web browsing, the C2-Ultra with two SME2 units completes the workflow 24 percent faster than the previous generation.
Bringing neural accelerators to mobile graphics #
Within CSS for Mobile 2, Mali G2-Ultra NX provides an AI-native graphics foundation for increasingly sophisticated neural graphics workloads. The GPU brings neural acceleration directly into the GPU, with neural accelerators tightly integrated into the shader cores, alongside a new execution engine and third-generation ray tracing unit. This complements the agentic AI capabilities of the C2 CPU cluster with dedicated GPU acceleration for increasingly demanding neural graphics workloads.
For more details about Mali G2-Ultra NX read this blog.
Heterogeneous platform compute through system IP #
Agentic AI and increasingly sophisticated graphics place greater demands on how compute and data move and how workloads communicate across a mobile system. CSS for Mobile 2 addresses this through the SI L2 system interconnect, which complements the CPU and GPU by helping to sustain responsive execution across AI and graphics workloads.
SI L2 provides lower-latency access, higher bandwidth across the platform, alongside coherency and quality-of-service capabilities that help manage and prioritize activities across multiple compute resources and heterogeneous workloads. These capabilities are particularly important as the CPU, GPU and other system resources operate concurrently and share access to data and memory.
Developer-ready from day one #
CSS for Mobile 2 is supported by a software ecosystem that helps more than 22 million developers building on Arm access its capabilities through familiar frameworks and development environments. Arm KleidiAI, framework integrations and new Arm AI Portal connect the CSS for Mobile 2 platform’s hardware capabilities with the models and software developers use to build increasingly capable on-device AI experiences.
KleidiAI and integrations with popular AI frameworks provide optimized software paths for Arm CPUs, including workloads that take advantage of SME2, so developers benefit from the underlying CPU capabilities through the software stacks they already use.
Arm AI Portal extends that support into AI application development, providing optimized and validated models, performance and accuracy data, code examples and deployment resources in one place, with the Arm MCP Server bringing these capabilities into agentic development environments. This helps developers identify the right model for their workload, understand how it performs on Arm and move quicker from evaluation to deployment.
Through AI Portal, developers targeting CSS for Mobile 2 can take advantage of Arm CPU capabilities, such as SME2-based AI acceleration on the C2 CPUs, and integrated neural acceleration on Mali G2-Ultra NX, connecting optimized AI models with the performance available on mobile devices.
The complete platform that brings agentic AI on device #
The next era of mobile AI will be decided by how well the whole system turns intent into action. CSS for Mobile 2 brings together best-in-class CPU and GPU, system IP, physical implementations and software in a configurable platform optimized for responsiveness, efficiency and sustained performance.
As agentic AI becomes more deeply embedded in everyday experiences, this system-level approach gives silicon partners the flexibility to differentiate and developers a faster path to bring increasingly capable on-device AI experiences to life.
Arm CSS for Mobile 2
An AI-native mobile compute platform for responsive agentic AI, neural graphics and sustained mobile performance.
Any re-use permitted for informational and non-commercial or personal use only.