Arm’s CSS for Mobile 2 Adds Dedicated Silicon for On-Device AI Agents and Neural Graphics

Arm has unveiled Compute Subsystem (CSS) for Mobile 2, its next reference platform for flagship smartphone chips, built around a new CPU cluster and what the company calls its first AI-native mobile GPU. The announcement came today at the Arm Everywhere China event in Shanghai and succeeds last year’s Lumex CSS platform, dropping the Lumex name entirely just a year after it replaced the long-running Cortex branding for Arm’s mobile CPU designs.

The pitch is that phones need dedicated silicon for agentic AI, not just faster general-purpose cores. Arm’s Executive Vice President of Edge AI, Chris Bergey, told The Register that AI has moved from background features like speech recognition into tools people actively direct, and that agentic AI goes further still: interpreting a goal, coordinating multiple applications and models, and carrying out a sequence of actions on its own. “So you have repeated inference, application logic, system orchestration, multiple workloads operating concurrently, and so that means you don’t just need more CPU performance,” Bergey said. “You need multiple high-performance CPUs working in parallel to keep all those different workloads running.”

That reasoning also underpins Arm’s case for keeping more AI processing on the device rather than routing it through the cloud. “When every interaction has to travel to a datacenter and back, there’s implications for latency, connectivity, the cost of delivering those services at enormous scale,” Bergey said, adding that he expects most workloads that can move to the network edge eventually will.

A CPU cluster built to run several agents at once

The new Arm C2 CPU cluster pairs a top-end C2-Ultra core, which Arm says delivers up to 15 percent higher single-thread performance and up to 1.7 times higher AI performance than last year’s C1-Ultra, with smaller, more efficient C2-Pro cores. An example flagship configuration combines two C2-Ultra cores with six C2-Pro cores and two Scalable Matrix Extension 2 (SME2) units, twice the SME2 capacity of the previous generation. Arm says doubling that matrix-math hardware delivers up to a 70 percent speedup running the latest small language models directly on the phone. Notebookcheck separately reported a 12 percent gain in multi-threaded performance generally, and that a specific two-C2-Ultra, two-SME2 configuration completed a full workflow 24 percent faster than the prior generation.

Arm frames the CPU cluster itself as the orchestration layer for an on-device agent: it maintains context and task state while deciding whether each step should run through a local app, an on-device model, another accelerator, or a cloud service. Higher single-thread performance is meant to cut the latency of individual steps, while the added multi-thread headroom lets retrieval, inference, and application logic run at the same time. Arm’s DynamIQ Shared Unit ties the cores in the cluster together, which the company says has been optimized to keep several concurrent agentic workloads responsive, while a separate system-level interconnect, the SI L2, provides coherency and quality-of-service-aware communication across the platform’s compute resources.

A GPU with neural acceleration built into the pipeline

On the graphics side, the Mali G2-Ultra NX is what Arm calls its first Mali design with dedicated neural accelerator hardware built directly into the shader cores rather than added as a separate block, so neural and conventional rendering work share the same memory system and caches. Arm also redesigned the GPU’s execution engine, which it describes as the largest change to the Mali instruction set in seven generations, giving complex shaders roughly twice as many registers to work with, and added a third-generation ray tracing unit with hardware support for Opacity Micromaps, which handles complex transparent geometry without the usual ray-tracing performance cost.

Three neural rendering techniques sit on top of that hardware: Neural Super Sampling (NSS), which reconstructs a 1080p image from a 540p render; Neural Frame Rate Upscaling (NFRU), which generates intermediate frames to raise the displayed frame rate; and Neural Super Sampling and Denoising (NSSD), which combines upscaling with ray reconstruction and denoising for heavier ray-traced scenes. Arm says combining NSS and NFRU lets the GPU render just one eighth of the pixels that ultimately appear on screen, with the neural hardware reconstructing the rest, an approach it says makes desktop-class graphics possible within a 1-watt power budget. In a separate demo Arm calls Neural Dawn, a combination of NFRU and NSSD delivered up to four times higher performance efficiency and cut external memory traffic by as much as 70 percent compared with native rendering, a distinct result from the NSS-plus-NFRU pixel-count figure above. Arm’s own published numbers for the chip more broadly claim up to 14 percent higher performance on existing, non-AI game content compared with the previous generation.

Arm’s GPU Product Manager, Deyan Lazarov, described a third scenario to The Register, this one specific to heavy ray tracing: “The G2-Ultra NX GPU is able to sustain a 30 frame per second gaming session on content with very rich and intricate ray traced scenery,” a result he attributed to a 70 percent cut in the underlying ray tracing workload that converts into a 30 percent performance boost on that kind of scene, a separate figure again from both the Neural Dawn results and the general non-AI gaming number.

Ecosystem and rollout

Arm named an initial ecosystem that includes Alipay, Google’s AI Edge Gallery, OPPO, vivo, Tencent Games, Unity China, NetEase, and Infold Games. NetEase’s game “Where Winds Meet” is due to ship an NSS-enabled version later this year, and Arm’s demo lineup included “Arena Breakout Infinite” and “Infinity Nikki.” Separately, Arm told The Register it also demonstrated other games running at 60 frames per second on the new platform, and it is offering Premium and Pro variants of the Mali G2 GPU so chipmakers can choose the performance tier that fits their design.

Devices built on CSS for Mobile 2 are expected to start showing up within weeks, according to Notebookcheck, with MediaTek first to market and Samsung’s Exynos 2700 following in early 2028. Because Arm licenses the platform’s components separately, chipmakers aren’t required to take the whole package: Notebookcheck noted that, as an example of that flexibility, Samsung could pair the new CPU cluster with its own in-house GPU design instead of Mali, and Google could do the same with GPU IP from Imagination Technologies. Notebookcheck reported that Qualcomm is expected to detail its own next-generation Snapdragon designs later this month, with Apple’s A20 Pro expected around its usual fall iPhone launch, setting up a direct comparison between rival approaches to on-device agentic AI.