Arm Targets the AI Smartphone Era With 'CSS for Mobile 2'

Arm unveiled its next-generation mobile computing platform, 'Arm CSS for Mobile 2,' on the 8th. 'Arm CSS for Mobile 2' was designed to meet the system-level requirements of agentic AI and next-generation cinematic mobile graphics.

The platform combines a GPU and CPU cluster with system technologies and a software foundation. Its components include the 'Arm Mali G2-Ultra NX' GPU, the 'Arm C2 CPU' cluster, enhanced system IP, physical implementation technology, and the software ecosystem.

Arm said the platform is intended to help silicon partners accelerate the launch of differentiated mobile products. It also aims to make it easier for developers to build and deploy AI workloads on a common platform.

The core of CSS for Mobile 2 is the direct integration of neural processing capabilities into the mobile GPU.

In mobile graphics, limits have recently emerged in improving both image quality and performance with conventional rendering alone. As a result, the use of neural network technologies is expanding, and there is a growing trend toward reconstructing and enhancing visual details. In gaming, demand is also rising for realistic lighting, high-resolution graphics, and smooth frame rates.

Against this backdrop, Arm introduced the 'Mali G2-Ultra NX' product. It is equipped with the first Mali GPU and a dedicated neural accelerator, and is designed to perform AI-based tasks inside the graphics pipeline based on tighter integration of conventional graphics processing and neural processing. Arm presented Mali G2-Ultra NX as a product that supports AI-native graphics at the mobile GPU level.

Arm said that when Mali G2-Ultra NX is combined with Arm neural network technology, neural graphics performance reaches up to 4 times higher per watt, while conventional graphics performance improves by up to 14% from the previous generation. It also said the product addresses increasingly complex graphics workloads and the latest game engine features by applying a new execution engine and a next-generation ray tracing unit (RTU).

Arm showcased examples of applying its neural graphics technology in game production through multiple engines and studios. It presented 'Neural Dawn,' co-developed by Arm and Sumo Digital, along with an AI graphics experience based on 'Neural Dawn' and Mali G2-Ultra NX. It also highlighted the integration of Tencent Games Central Tech's 'Magic Dawn' and Arm neural graphics technology into developer workflows, as well as Unity China's 'Tuanjie Engine' and Arm neural graphics technology developer workflow integration. Integration of Arm neural graphics technology into major game engines and studios is also under way.

At the same time, technology applications and demonstrations such as NSS and NSSD are continuing in individual games. NetEase Games' 'Where Winds Meet' is scheduled to release a version with NSS applied this year. A demonstration of NSSD technology using Tencent Games' 'Arena Breakout Infinite' was also conducted. NSS integration for Infold Games' 'Infinity Nikki' is also under way.

Separately, in the CPU area, the importance of execution and orchestration has come to the fore as AI agents take on context retention, app execution, and coordination of multiple models and services. The CPU segment's focus has shifted to the execution and orchestration of agentic AI. AI agent capabilities are expanding beyond running a single AI model to maintaining user context, launching applications, and coordinating multiple models and services. As demand for agentic AI grows, the importance of the CPU's role is also increasing.

Arm unveiled a new 'C2 CPU cluster' designed to meet AI processing demands, saying it combines the C2-Ultra CPU, the C2-Pro CPU, and two SME2 units. Arm said the C2-Ultra CPU is built for peak performance, while the C2-Pro CPU emphasizes efficiency. It also said the C2 CPU cluster supports low-latency AI processing directly on the CPU and integration with AI accelerators.

Arm said SME2 performance within the cluster has doubled, and processing speed for the latest small language models (SLMs) has improved by up to 70%. It also said the C2-Ultra delivers up to 1.7 times better AI performance than the previous-generation C1-Ultra, 15% better single-thread performance, and up to 38% lower power consumption at the same performance level. Arm also said the SME2 ecosystem is expanding, announcing that Alipay, Google AI Edge Gallery, OPPO, and vivo are participating in the SME2 ecosystem, and added that related technology is being applied to major Android and iOS smartphones.

Arm is strengthening its software ecosystem to improve utilization of its new hardware. To that end, Arm introduced CSS for Mobile 2 based on KleidiAI and the Arm Neural Graphics Development Kit, and said CSS for Mobile 2 helps developers make use of the new hardware capabilities. Arm also unveiled a new public 'AI Portal' as a unified touchpoint for developers and AI agents.

CSS for Mobile 2 and AI Portal are pillars that help developers combine and use multiple compute resources. AI Portal provides search, evaluation, and utilization functions for optimized models, code, and tools. Arm's strategy is to build a mobile AI environment based on using a single optimized system of diverse computing resources such as the CPU, GPU, and neural accelerators, rather than relying on specific accelerators.

This direction also aligns with changes in mobile AI. AI is expanding from smartphone feature-level functions to agentic AI that continuously understands user needs and carries out tasks. Accordingly, Mobile SoCs are also moving away from an approach that strengthens the CPU, GPU, and NPU individually and toward improving the overall system's AI processing efficiency and responsiveness.

Arm's CSS for Mobile 2 is a platform designed to respond to recent changes, targeting both agentic AI and AI-native graphics. Because of these characteristics, it is expected to influence future competition in mobile SoC design as well.

Source: TECHWORLD · Park Gyu-chan
Original: https://www.epnc.co.kr/news/articleView.html?idxno=406660

References

• https://www.epnc.co.kr/news/articleView.html?idxno=406660