On September 8, at its Arm Everywhere China annual event in Shanghai, Arm unveiled the second-generation Compute Subsystem for Mobile—CSS for Mobile 2. The platform integrates the new C2 CPU cluster, the Mali G2-Ultra NX GPU, enhanced system IP, and a developer-ready software ecosystem.
The shift is not just about faster chips. It's about rethinking how mobile devices handle intelligent agents that perceive, remember, reason, and act continuously. Arm framed the release around one question: how to make a phone smart enough to manage a multi-step task like dinner reservations—without burning through its battery in an hour.
C2 CPU: Orchestrator, Not Calculator
Arm's C2 CPU cluster marks a strategic shift. It no longer positions the CPU as just a compute engine. It's now the "orchestration layer" for agentic AI workflows.
The cluster combines the performance-focused C2-Ultra core, the efficiency-optimized C2-Pro core, and dual SME2 units. SME2 is the Armv9 matrix math extension that enables matrix operations directly on the CPU—bypassing the need to shuttle data to a separate NPU for every AI task.
Key performance gains over the previous C1 generation:
AI model performance up 1.7x
Single-thread performance up 15%
End-to-end agentic workflow performance up 24%
Speech-to-text model latency down 40%
Up to 38% power reduction at same performance
The "orchestration" framing matters. In an agentic workflow—perceiving context, retrieving memory, reasoning, executing actions—the CPU must manage task state, coordinate resources, and decide whether to use local apps, on-device models, or cloud services. The C2 cluster is designed to do that without the system stalling.
Arm's reasoning: Google Play's AI apps run 95% on CPU. Better CPU AI acceleration directly improves the majority of on-device AI experiences.
The First GPU with a Neural Accelerator Inside
The Mali G2-Ultra NX is the first Mali GPU to integrate a dedicated neural accelerator directly into the shader cores. It's not a separate NPU bolted on—it's embedded in the graphics pipeline itself.
This enables three neural graphics technologies:
NSS (Neural Super Sampling): Reconstructs high-quality frames from low-resolution renders—similar to desktop DLSS
NFRU (Neural Frame Rate Upscaling): Generates intermediate frames via AI for smoother gameplay
NSSD (Neural Super Sampling with Denoising): Combines upscaling and denoising for ray-traced scenes
Performance claims are aggressive. In neural graphics workloads, the GPU delivers up to 4x better per-watt performance with 70% less DRAM traffic. Traditional gaming performance is up 14% over the previous generation.
The neural accelerator runs at 1-2W for neural graphics—a realistic power budget for mobile. It supports INT8/INT16 inference and handles compressed weights to reduce memory bandwidth.
Arm Just Wrote the Mobile AI Playbook
Arm's roadmap is clear. In the AI-native era, performance is not defined by a single component. It's defined by the system's ability to convert user intent into action, across the full compute stack—CPU, GPU, NPU, and memory.
CSS for Mobile 2 brings that vision to production. The C2 cluster handles orchestration and low-latency AI inference directly on the CPU. The Mali G2-Ultra NX handles neural graphics and visual AI directly in the GPU pipeline. The SI L2 system interconnect reduces CPU-to-DRAM latency by about 45%, keeping data moving without bottlenecks.
The ecosystem is already moving. NetEase's Where Winds Meet will be among the first games to support NSS. Tencent's Magic Dawn and Unity China's Tuanjie Engine have integrated Arm's neural technology. Xiaomi's recently announced Xuanjie O3 chip uses a 16-core G2-Ultra NX GPU.
Arm is not building the chip. But it's building the blueprint for everyone who does.
P.S. The quiet detail in the announcement: Arm explicitly positioned CSS for Mobile 2 for "agentic AI," not just "AI features." The distinction is the difference between a chatbot and an agent that does real work. Every major mobile SoC design runs on Arm. If Arm's blueprint works, the next generation of phones won't just run AI—they'll run agents. And they'll do it without burning through your battery.
Frequently Asked Questions
Q: What is Arm CSS for Mobile 2?
A: Arm CSS for Mobile 2 is the second-generation Compute Subsystem for Mobile, announced on September 8, 2026. It integrates the C2 CPU cluster, Mali G2-Ultra NX GPU, enhanced system IP, and software ecosystem for AI-native mobile computing.
Q: What is the C2 CPU cluster?
A: The C2 CPU cluster is Arm's new CPU design for agentic AI workflows. It features dual SME2 units for matrix math acceleration, with AI model performance up 1.7x, single-thread performance up 15%, and end-to-end agentic workflow performance up 24% over the previous generation.
Q: What is SME2?
A: SME2 (Scalable Matrix Extension version 2) is Arm's matrix math extension that enables matrix operations directly on the CPU—bypassing the need to shuttle data to a separate NPU for every AI task. It doubles SME2 compute capacity compared to the previous generation.
Q: What is the Mali G2-Ultra NX?
A: The Mali G2-Ultra NX is the first Mali GPU with a dedicated neural accelerator built directly into the shader cores. It enables neural super sampling, frame rate upscaling, and denoising directly on the GPU.
Q: What are the neural graphics technologies?
A: Three technologies: NSS (Neural Super Sampling) for high-quality upscaling, NFRU (Neural Frame Rate Upscaling) for AI-generated intermediate frames, and NSSD (Neural Super Sampling with Denoising) for ray-traced scenes.
Q: How much performance gain does the Mali G2-Ultra NX deliver?
A: In neural graphics workloads, the GPU delivers up to 4x better per-watt performance with 70% less DRAM traffic. Traditional gaming performance is up 14% over the previous generation.
Q: Which games are adopting the technology?
A: NetEase's Where Winds Meet will be among the first games to support NSS. Tencent's Magic Dawn and Unity China's Tuanjie Engine have integrated Arm's neural technology.
Q: What is Arm's strategy with CSS for Mobile 2?
A: Arm is positioning CSS for Mobile 2 for "agentic AI"—not just AI features. The platform rethinks how mobile devices handle agents that perceive, remember, reason, and act continuously, with the CPU as the "orchestration layer" for agentic workflows.
Q: What is the significance of "agentic AI" in this context?
A: Agentic AI refers to intelligent agents that can perform multi-step tasks autonomously—like making dinner reservations or managing complex workflows—rather than just responding to single prompts. Arm is designing hardware specifically for this use case.
Q: Is Arm building the chip itself?
A: No. Arm builds the blueprint (IP) that chipmakers like Qualcomm, Samsung, MediaTek, and Xiaomi license to build their own chips. CSS for Mobile 2 is the reference design that these chipmakers will adopt for their next-generation mobile SoCs.
