Arm targets the agentic AI era with new edge, cloud and robotics platforms

Arm is expanding its computing platform for an era in which artificial intelligence does more than respond to individual prompts. The company expects AI agents to operate continuously across phones, data centers, vehicles, robots and other physical systems—perceiving their surroundings, reasoning about objectives and taking actions on a user’s behalf.

Three announcements underpin that strategy: the Arm CSS for Mobile 2 platform for edge devices, Neoverse CSS N4 for cloud infrastructure, and an Arm Total Design initiative for physical AI. Together, they aim to provide a common architecture and software foundation across a highly fragmented computing landscape.

At the edge, Arm is introducing a platform combining its C2-Ultra CPU cluster with the Mali G2-Ultra NX GPU. The C2-Ultra incorporates SME2, the second generation of Arm’s Scalable Matrix Extension. SME2 accelerates the matrix operations used heavily by machine-learning models while retaining the flexibility of a general-purpose CPU.

Arm claims that two SME2-enabled C2-Ultra cores deliver 1.7 times the AI performance of its previous design across recent models. In a representative flagship configuration, the company also reports 15% higher single-thread performance, 15% faster web browsing, 12% faster application launches and 12% greater multithread performance.

That CPU improvement matters because an AI agent is not simply a neural-network inference workload. A useful agent must retrieve context, run applications, call tools, access databases, communicate over a network and verify its results. Much of that work depends on CPUs and operating-system services. Faster local execution can reduce the delay between a request and an agent’s action while keeping sensitive personal context on the device.

The Mali G2-Ultra NX addresses a different bottleneck: the escalating cost of high-quality graphics. Its neural accelerators are integrated directly into the GPU’s shader cores and support Neural Super Sampling, Neural Frame Rate Upscaling, and combined super-sampling and denoising. These techniques render less information conventionally and use learned models to reconstruct higher-resolution images or additional frames.

Arm says its neural upscaling can convert 540p output to 1080p, while frame-rate upscaling can turn 30 frames per second into 60. The company claims up to a fourfold increase in frame rate and efficiency, alongside a 70% reduction in DRAM traffic. A redesigned execution engine and third-generation ray-tracing unit target more conventional rendering, including Unreal Engine 5’s Nanite and Lumen technologies. Arm reports a 24% benchmark uplift, a 14% gain in non-AI gaming and a 70% reduction in ray-tracing workload.

In data centers, Neoverse CSS N4 packages CPU cores, cache, memory management, interconnect and system IP into a configurable compute subsystem. Customers can adjust core count, cache capacity, I/O and connectivity while avoiding much of the integration work required to assemble custom silicon. Configurations scale to 128 cores per die. Compared with Neoverse CSS N3, Arm claims twice the socket performance, 1.25 times the performance per watt and 1.75 times the memory bandwidth under its stated reference conditions.

The technical significance is configurability. Agentic workloads mix model inference with CPU-intensive orchestration, sandbox execution, storage, networking and databases. Cloud providers therefore need both dense, efficient infrastructure and higher-performance processors. CSS N4 is intended to shorten the path to chips optimized for those differing requirements.

Arm’s physical-AI initiative extends the platform into robotics, industrial machinery and vehicles, where “photon-to-torque” latency—the time between sensing the environment and producing physical movement—is critical. Arm Total Design for Physical AI will bring silicon vendors, model developers, sensor suppliers, manufacturers and safety specialists together. A proposed Robotics Capability Framework would give those participants a common way to describe reusable capabilities.

Bottom line: Arm is positioning itself as connective tissue for distributed AI. Its advantage is not one accelerator or model, but an architecture already spanning billions of devices and supported by more than 22 million developers. If Arm can make software and AI capabilities portable from cloud systems to phones and machines, developers could build agents once and deploy each component where latency, privacy, power consumption and performance make the most sense.

Also Read:

Security by Design: Arteris Expands Its Partnership with Arm

Arm FCSA and the Journey to Standardizing Open Chiplet-Based Design

Arm Lumex Pushes Further into Standalone GenAI on Mobile