Arm is stepping up efforts to tie smartphones, robots and data centres into a single computing platform. Arm held a media briefing on Sept. 7 and unveiled an AI computing portfolio. It introduced three products: CSS for Mobile 2, a new mobile compute subsystem (CSS); Arm Total Design for Physical AI targeting robots and autonomous devices; and the Neoverse CSS-N4 for data centres.

The products Arm unveiled share a common target: the computing load created by agentic AI. AI agents do more than run models. They call tools, access databases to retrieve information and coordinate other agents. Arm said that way of working has increased CPU load. As tasks are carried out on the CPU, a single user request can expand into simultaneous work by multiple agents, adding to the burden on the CPU.

CSS for Mobile 2 is designed to address this. It is a compute subsystem that chip designers can use as a complete package rather than combining CPUs and GPUs separately. Arm stressed it is a mobile platform designed on the assumption of AI use. Arm said it improved single-thread performance of the C2 CPU cluster by 15 percent and multi-thread performance by 12 percent from the previous generation.

Arm said it enabled up to 2 SME2 units for AI computation inside the CPU to be included in the C2 cluster. Arm's SME (Scalable Matrix Extension) is an extension that adds matrix operations to the CPU instruction set. It is a unit that handles AI computation within the CPU, allowing matrix operations to run directly on the CPU without going through a separate NPU.

James McNiven (제임스 맥니븐), vice president of client computing and edge AI at Arm, said, "Based on 2 SME2 units, we expect about 5 to 7 TOPS." SME2 is already included in major Android and iOS devices. Arm said 95 percent of AI applications on the Google Play Store run on the CPU.

Arm also changed the GPU rendering approach. The new Mali G2 places a neural-network accelerator inside the shader to handle neural-network super-sampling that boosts images rendered at lower resolution. Arm said frame performance rose fourfold from the previous generation within the same power range, and DRAM power consumption also fell. It also said ray tracing, which developers had used sparingly due to power limits, can be increased to about 1,000 rays.

G2 Ultra NX includes at least 6 neural-network accelerators, and Arm projected it would typically be used in the 10 to 20 TOPS range. Cumulative shipments of Mali GPUs total 14 billion units. Arm said it has completed adoption in major game engines and commercial titles, and the feature can be used from devices launched at the end of the year.

Arm also unveiled a physical AI programme targeting robots and autonomous devices. Arm Total Design for Physical AI includes more than 80 partners, including AWS, Hugging Face, NXP, Siemens and Unitree. Dermot O'Driscoll (더못 오드리스콜), vice president for physical AI go-to-market at Arm, forecast that the physical AI computing market, which was $25 billion last year, would exceed $200 billion in the 2030s. He said in this area, the latency from when sensors capture signals to when actuators move, and the weight of computing devices, determine design.

On the data centre side, Arm introduced Neoverse CSS-N4. Socket-level performance is more than twice that of its predecessor, CSS-N3. Performance per watt increased 25 percent, and memory bandwidth rose 75 percent. Arm said it has supplied 1.5 billion cores to data centres since introducing Neoverse in late 2018, and 500 million of those cores were shipped over the past 9 months this year. Arm pointed to the ability for partners to adjust core and cache capacity and input-output configurations as a feature of the product.

Samsung Electronics was also mentioned in an explanation of manufacturing processes. Eddie Ramirez (에디 라미레즈), vice president for cloud AI marketing at Arm, said the main design points for CSS-N4 are TSMC N3 and N3P, adding, "Samsung Electronics' SF2 and SF2P are also design points that are in high demand."

Arm benchmarked CSS for Mobile 2 based on TSMC N2. McNiven said, "This platform can be ported to multiple foundries and processes," and added, "Process choices by partners will become clear within a few weeks."