Arm has introduced the CSS for Mobile 2 computing subsystem, a platform designed for agentic AI and AI-native graphics. The system integrates the Arm C2 CPU cluster, Mali G2-Ultra NX GPU, and SI L2 system interconnect, combining system IP, software, and developer tools into a a single platform.
The Arm C2 CPU cluster features the C2-Ultra and C2-Pro CPUs along with second-generation Scalable Matrix Extension (SME2) technology. This cluster delivers up to 1.7x performance gains in the latest AI models, a 15% increase in single-thread performance, and 15% faster web browsing. Application launch speeds have improved by 12%, and cluster-level multi-thread performance has risen by 12%.
In agentic AI workflows, the C2-Ultra CPU reduces latency by 40% when running Moonshine and Parakeet speech-to-text models compared to the C1-Ultra. The average speed for generating structured instructions using small language models with retrieved context has increased by 25%. Overall end-to-end performance for tasks including voice processing and memory retrieval has improved by 24%.
The Mali G2-Ultra NX is Arm's first AI-native Mali GPU, featuring neural network accelerators embedded directly into the shader cores. It includes a third-generation ray tracing unit (RTUv3) and a new execution engine. This architecture supports Neural Super Sampling (NSS), Neural Frame Rate Up (NFRU), and Neural Super Sampling and Denoising (NSSD).
In a demo game developed with Sumo Digital using Unreal Engine, NFRU and NSSD solutions achieved up to 4x performance efficiency gains and reduced external memory traffic by up to 70%. The GPU supports stable operation at up to 120 fps. Benchmark performance has increased by up to 24%, and non-AI gaming performance has risen by 14%.
The platform is supported by a developer community of over 22 million people. Arm provides the KleidiAI integration for AI frameworks and the Arm AI Portal for optimized models and deployment resources. The Arm MCP server brings these capabilities into agentic AI development environments to accelerate the deployment process.
