Arm's new Mali G2-Ultra NX GPU adds neural accelerators for sub-1-watt AI graphics on mobile

Arm announced a series of agentic AI-focused enhancements to its IP offerings at its Arm Everywhere China event. The news encompasses mobile devices, physical AI and robotics, and AI infrastructure for cloud and enterprise environments. The semiconductor industry has seen its rate of change increase dramatically, driven by the transition to agentic AI workloads and new demands on computing devices, both on the edge and in the datacenter.

CSS for Mobile 2 and the Mali G2-Ultra NX

Arm highlighted efforts to make working with its offerings easier, from its CSS (Compute Subsystem) concept to its Total Design principles. CSS allows companies to start with a fully designed platform, while Total Design pulls together multiple partners across an ecosystem. The company also emphasized the range and consistency of its development tools and announced a new Arm AI portal that aggregates tools, models, and agentic harnesses optimized for the Arm platform. Arm has a developer base of 22 million people creating software for billions of Arm-based devices across numerous product categories.

CSS for Mobile 2 brings CPU, GPU, and system technologies together in an AI-native compute platform for edge devices. Individual chip architecture enhancements include:

• Mali G2-Ultra NX - a new GPU design with AI-enabling neural accelerators for the first time, enabling AI-generated graphics at sub-1-watt power levels
• C2-Ultra and C2-Pro CPUs
• Two blocks of compute designed to accelerate Arm's SME2 instruction set extensions, which are optimized for matrix math in agentic AI workloads and can bring up to 1.7x improvements

NSS reconstructs higher-resolution images from lower-resolution inputs, combining temporal data and integrated anti-aliasing to improve image quality efficiently on mobile.

Physical AI and Robotics

Arm brought together over 80 companies under Arm Total Design for Physical AI, patterned after the existing Total Design for Cloud AI. The group spans sensor and actuator companies, operating system vendors, software development firms, and semiconductor partners.

Arm also released a Robotics Capability Framework that defines six types of autonomous robots with increasing capabilities and self-governance, loosely based on the five levels of autonomous driving.

AI Infrastructure

Arm delivered Neoverse CSS N4, which adds support for designs with up to 128 cores, LPDDR6 memory, and PCIe Gen 7 connectivity. The company also highlighted early success with its own AGI CPU chip, introduced at the first Arm Everywhere event earlier this year.

Bob O'Donnell is the founder and chief analyst of TECHnalysis Research, LLC, a technology consulting firm that provides strategic consulting and market research services to the technology industry and professional financial community. You can follow him on Twitter @bobodtech.