Mobile chips have spent years chasing benchmark numbers. Arm is now betting the next race is about something harder: keeping an AI agent context-aware, responsive, and running within a smartphone’s thermal limits, all while rendering console-quality graphics. That’s a genuinely difficult system-level problem, and it’s what CSS for Mobile 2 is designed to solve.
Arm announced CSS for Mobile 2 on September 8, 2026, packaging together a new GPU, a new CPU cluster, updated system IP, and a developer software stack into a single reference platform for next-generation mobile silicon.
The GPU: neural accelerators built into the graphics pipeline
The Mali G2-Ultra NX is the headline. It’s the first Mali GPU to include dedicated neural accelerators, tightly coupled with traditional rasterization so graphics and neural workloads share the same pipeline. Arm claims up to 4x higher performance per watt for neural graphics tasks compared to the previous generation, which is a significant jump if it holds in real workloads. There’s also a new execution engine and an updated Ray Tracing Unit. For traditional game content, Arm says performance is up 14% over the prior generation. Developers already working with this include Tencent Games, Unity China’s Tuanjie Engine, and NetEase, whose title ‘Where Winds Meet’ is planning an NSS-enabled release this year.
The CPU cluster: SME2 doubles down on AI
The C2 CPU cluster pairs the C2-Ultra, Arm’s highest-performance mobile CPU, with efficiency-focused C2-Pro cores and two SME2 units. That doubled SME2 configuration is supposed to deliver a 70% speedup on small language models. Compared to the C1-Ultra, the C2-Ultra offers 1.7x higher AI performance, 15% better single-thread performance, and up to 38% lower power at equivalent performance levels. So it’s faster and cheaper to run. SME2 already has real ecosystem traction, appearing in leading Android and iOS handsets with support from Alipay, Google AI Edge Gallery, OPPO, and vivo.
Why this matters beyond the specs
Qualcomm’s Snapdragon 8 Elite and Apple’s A18 Pro are the obvious comparisons here. Both integrate NPUs tightly with CPU and GPU, but CSS for Mobile 2 is Arm’s attempt to give its silicon partners a common reference platform that bakes in the AI-native assumptions from the start, rather than bolting on inference engines as an afterthought. For chip designers licensing from Arm, this shortens the path to a differentiated product. For developers, the new AI Portal and updated KleidiAI tooling aim to make targeting these capabilities less painful. That’s the part worth watching: hardware is only as useful as the software that can actually reach it.
