Arm has introduced CSS for Mobile 2, a new compute platform for next-generation mobile devices. It brings together agentic AI and AI-native graphics to support the next generation of AI workloads on smartphones.
The platform is designed for AI workloads that require devices to maintain context, run applications, coordinate models and services, create content and act on behalf of users, while operating within mobile power and thermal limits.
Mali G2-Ultra NX brings neural acceleration to graphics
The Mali G2-Ultra NX is the first Mali GPU with dedicated neural accelerators, combining neural and traditional graphics processing in the same pipeline. It features a new execution engine and a next-generation Ray Tracing Unit for increasingly complex graphics workloads and modern game engine features.
With Arm Neural Technology, the GPU provides up to 4x higher performance per watt for neural graphics. The technology can reconstruct, enhance and generate visual detail, while Arm says neural graphics can deliver up to 4x the frame rate at a sustained power budget compared with native rendering. For traditional graphics, the GPU offers up to 14% higher performance on existing game content compared with the previous generation.
Arm highlighted several implementations across games and game engines:
- Neural Dawn, developed with Sumo Digital, demonstrates the new graphics capabilities.
- Tencent Games Central Tech’s MagicDawn and Unity China’s Tuanjie Engine are integrating Arm Neural Technology.
- NetEase plans to bring an NSS-enabled version of Where Winds Meet to players this year.
- Tencent Games’ Arena Breakout Infinite is being used for an NSSD technology demonstration.
- Infold Games’ Infinity Nikki has integrated NSS.
C2 CPU cluster targets agentic AI
The Arm C2 CPU cluster combines the high-performance C2-Ultra and efficiency-focused C2-Pro CPUs with two SME2 units. It is designed to orchestrate agentic AI workloads by maintaining context, scheduling workloads and coordinating processing across the compute system as AI agents reason, plan and act.
The additional SME2 capability supports responsive, low-latency AI processing directly on the CPU while working alongside other accelerators. Arm says the doubled SME2 capability delivers up to a 70% speedup on the latest Small Language Models.
Compared with the C1-Ultra, the C2-Ultra provides:
- Up to 1.7x higher AI performance
- 15% higher single-thread performance
- Up to 38% lower power consumption at the same performance
Separately, Arm says CSS for Mobile 2 enables up to 1.7x faster execution for AI models supporting voice, personal memory and reasoning. SME2 is already present in leading Android and iOS handsets, with the ecosystem including Alipay, Google AI Edge Gallery, OPPO and vivo.
System and software support
The platform includes the SI L2 system interconnect, designed to provide lower latency and higher bandwidth for system-level AI performance. Its configurable design allows silicon partners and OEMs to tailor performance, power and product requirements when developing custom SoCs, with physical implementations included alongside the compute IP.
Arm’s software and developer support includes:
- Arm KleidiAI
- Arm Neural Graphics Development Kit
- Software and open interfaces for the platform’s capabilities
- AI Portal, a single entry point for developers and agents to discover, evaluate and use optimized models, code and tools
Quick Specs
- Platform: Arm CSS for Mobile 2
- CPU: Arm C2 CPU cluster
- CPU variants: C2-Ultra, C2-Pro
- AI acceleration: 2× SME2 units
- GPU: Arm Mali G2-Ultra NX
- Neural acceleration: Dedicated neural accelerators
- Ray tracing: Next-generation Ray Tracing Unit
- System interconnect: SI L2
- Neural graphics: Arm Neural Technology
- Software: Arm KleidiAI, Arm Neural Graphics Development Kit
- Developer platform: AI Portal
- Target devices: Next-generation mobile devices
On-device AI applications
CSS for Mobile 2 supports AI workloads that can understand user intent, retain context, coordinate tasks, create content and act on behalf of users. Agentic assistants can remember context and complete tasks across applications, with the C2 CPU cluster and SME2 providing responsive on-device compute.
For live translation, C2 with SME2 supports more accurate, low-latency language processing directly on the device. The platform also supports multimodal AI workloads that combine voice, images, text and context by coordinating CPU, GPU and system resources.
For AI-native graphics, the Mali G2-Ultra NX uses Arm Neural Technology to reconstruct and enhance visual detail while operating within mobile power limits.
Arm’s ecosystem partnerships
CSS for Mobile 2 is designed for an ecosystem spanning silicon partners, OEMs, game studios, game engines and AI developers. This includes work on neural graphics using Arm Neural Technology and Mali G2-Ultra NX, as well as on-device agentic AI.
Arm is working with Google on heterogeneous computing to run latency-sensitive AI workloads locally, including voice assistants, live translation, speech recognition, multimodal AI and personalized agentic experiences.