Arm introduces new AI-native compute platform built for agentic AI and mobile graphics
Arm has introduced CSS for Mobile 2, a compute platform designed for smartphones that increasingly need to run persistent AI agents alongside advanced graphics. The platform brings together a new GPU with dedicated neural accelerators, a high-performance CPU cluster with expanded AI capabilities, supporting system IP, physical implementations and software tools for developers. as reported by Newsroom Arm The announcement reflects a shift in how Arm defines mobile workloads. An agent running on a phone must do more than execute a single inference request: it may need to preserve context, launch applications, coordinate models and services, and act for the user while operating within tight power and thermal limits. CSS for Mobile 2 is intended to address those system-level requirements rather than treating AI as the job of one standalone accelerator.
A GPU designed for neural graphics
The centerpiece of the graphics platform is the Arm Mali G2-Ultra NX, which Arm describes as its most advanced mobile GPU and its first Mali GPU with dedicated neural accelerators. The design combines neural and traditional graphics processing in the same pipeline, allowing neural workloads to run alongside the graphics processing already taking place. The developer tools story also surfaces in Google launches Gemini 3.8 Flash and..., adding another angle.
That approach is intended to support techniques that reconstruct, enhance or generate visual detail instead of relying exclusively on conventional rendering. Arm says the GPU, used with Arm Neural Technology, delivers up to four times higher performance per watt for neural graphics. It also includes a new execution engine and a next-generation Ray Tracing Unit for more demanding game workloads, while offering up to 14% higher performance on existing game content compared with the prior generation.
Arm is positioning the technology for richer mobile games, including more detailed environments, cinematic lighting and higher-resolution gameplay. Its Neural Dawn demonstration, developed with Sumo Digital, showcases the intended direction. Arm also cites work involving Tencent Games Central Tech’s MagicDawn, Unity China’s Tuanjie Engine, and planned or demonstrated integrations for titles including Where Winds Meet, Arena Breakout Infinite and Infinity Nikki.
CPUs take on the agent’s orchestration work
CSS for Mobile 2 also introduces the Arm C2 CPU cluster. It combines the Arm C2-Ultra, described as Arm’s most powerful mobile CPU, with efficiency-focused C2-Pro CPUs and two SME2 units. The same ai agents question is explored in Prove2Me, which adds a research perspective.
Arm says the doubled SME2 capability enables a 70% speedup on the latest Small Language Models. Compared with the C1-Ultra, the C2-Ultra provides up to 1.7 times higher AI performance and 15% higher single-thread performance, while using up to 38% less power at the same performance.
The CPU’s role extends beyond raw model execution. In Arm’s platform design, it maintains context, schedules tasks and coordinates activity across the device’s CPUs, GPUs and other accelerators. That makes the CPU central to keeping agentic applications responsive as they reason, plan and act. The developer tools story also surfaces in Andrew Ng Launches OpenWorker to Deliver..., adding another angle.
Software support will determine adoption
Arm says CSS for Mobile 2 is intended to be developer-ready from launch, building on software including Arm KleidiAI and the Arm Neural Graphics Development Kit. The company is also introducing an AI Portal as a single entry point for discovering, evaluating and using optimized models, code and tools.
For silicon partners, the platform is designed to shorten the path to differentiated products. For developers, Arm is offering a shared foundation for building and deploying AI workloads across devices. The open question is how quickly handset makers and software teams will turn the combined CPU and GPU capabilities into everyday agentic features and neural-graphics experiences.