Arm, a British semiconductor design asset (IP) corporations, unveiled a next-generation mobile computing platform focused on running artificial intelligence (AI) directly on smartphones. As AI evolves from merely responding to user commands to an "agentic AI" that plans and performs multiple tasks on its own, the strategy is to boost the smartphone's on-device AI computing performance.
Arm said on the 7th that it would unveil its next-generation mobile platform, "Arm CSS for Mobile 2." CSS (Compute Subsystems) is a kind of "semiconductor design package" that bundles the key elements needed to design chips—central processing units (CPU), graphics processing units (GPU), and system IP—in advance. It lets companies that make smartphone chips reduce the effort of assembling each component from scratch and shorten new product development time.
The shift Arm is focusing on is AI computation moving from data centers down to smartphones. Currently, Generative AI handles a significant portion of complex computation in cloud data centers. But as AI agents begin searching for information on behalf of users, running applications, or collaborating with other AI, the computations that smartphones must handle themselves increase significantly.
Ami Badani, Arm chief marketing officer (CMO), said, "Agents call tools, access databases, search for information, and coordinate other agents," and added, "Considering token expense, privacy, and latency, it's not reasonable to send all interactions back to the cloud." Arm expects AI to be distributed so that large-scale AI models are handled in the cloud, while tasks requiring personal data or immediate responses are handled on devices close to users, such as smartphones.
To that end, Arm introduced the new "C2-Ultra" CPU and "Mali G2-Ultra NX" GPU. C2-Ultra targets the agentic AI environment in which multiple AI programs run simultaneously. Compared with the previous generation, single-thread performance is up 15% and multi-thread performance is up 12%. When using two cores with SME2 technology for fast AI computation, performance on the latest AI models improves by up to 1.7 times.
Chris Bergey, Arm executive vice president and general manager for Edge AI, said, "In the agent era, it's not just that users command 'do this'; multiple agents perform different tasks simultaneously in the background," explaining that CPU processing performance and scalability are becoming even more important. According to Arm, about 95% of AI applications registered on the Google Play Store use the CPU to handle AI computation.
Arm offered an example in which a user instructs a smartphone AI, "Plan and book an anniversary dinner." After understanding the user's intent, the AI searches for information, makes judgments, and then sequentially runs maps and reservation services. Arm said performing such tasks on C2-Ultra is about 15% faster than the previous generation, and about 24% faster when leveraging SME2.
Changes to the GPU targeting smartphone games are also significant. The Mali G2-Ultra NX is Arm's first "AI-native" Mali GPU. Instead of drawing every pixel on the screen at high resolution, it computes some at lower resolution and uses AI to fill in the details. It effectively brings AI upscaling technology used in PC graphics cards to smartphones, where power and heat constraints are tight.
For example, a screen rendered at 540p can be upscaled by AI to about 1080p quality, or a 30-frames-per-second screen can be increased to 60 frames. AI also removes noise generated during ray tracing, which realistically renders reflections and shadows of light. To do this, the GPU integrates a dedicated AI compute unit called the "Neural Accelerator (NX)."
Arm said that when using AI graphics features, frame performance and power efficiency each improve by up to four times compared with the previous generation, while data traffic to and from DRAM drops 70%. Not all GPU tasks are four times faster. Performance gains in regular games that don't use AI are 14%, and 24% on legacy-style graphics benchmarks.
Arm is expanding its AI computing reach beyond mobile to data centers and even robots and autonomous vehicles. For data centers, it unveiled "Neoverse CSS N4." It can configure up to 128 CPU cores on a single die, and compared with N3, performance per socket doubles, performance per watt rises 1.25 times, and memory bandwidth increases 1.75 times. The cumulative number of Neoverse cores shipped to data centers worldwide has surpassed 1.5 billion, with 500 million shipped in the past nine months.
It also presented "physical AI," in which AI perceives the real world and acts directly, such as in robots and autonomous vehicles, as a new growth pillar.
Drew Henry, Arm executive vice president for physical AI, described "latency" as the core of physical AI. Citing cars as an example, Henry said what matters is "the time it takes from the moment a camera sensor detects something to when the actual system operates," adding, "It's like how quickly a car driving on a highway applies the brakes after detecting an obstacle." In robots or drones, the weight of the computing device also matters. Henry said, "In physical AI, every 1 gram matters."
Arm projected that the computing market for physical AI will grow from about $25 billion last year to more than $200 billion annually in the 2030s. Henry said, "That's roughly 10 times growth in less than 10 years," adding, "Even this outlook may underestimate the market."
To target this, Arm is launching the "Arm Total Design for Physical AI" program, with participation from more than 80 corporations across semiconductors, sensors, cloud, and robotics. Like how the auto industry classifies autonomous driving levels from 1 to 5, it also plans to work with the industry to develop a "Robotics capability framework" that classifies robot performance and functions by common criteria.
Arm is touting as a competitive edge that the same Arm-based software ecosystem can be used across the cloud, smartphones, and robots and cars. CMO Badani said, "Everywhere AI compute is needed is already where Arm exists," adding, "AI is created in the cloud, personalized at the edge, and increasingly extends into the physical world." More than 22 million developers are currently participating in the Arm-based software ecosystem.