Arm has unveiled CSS for Mobile 2, a compute platform designed to help chipmakers build smartphones capable of running AI agents and richer graphics on-device.
The platform brings together the new Mali G2-Ultra NX GPU, C2 CPU cluster, enhanced system IP and software tools. For Arm licensees and device partners, the platform offers a coordinated foundation for developing mobile chips instead of integrating CPU, GPU, AI and system components separately.
The centerpiece is the Mali G2-Ultra NX, which Arm describes as its most advanced mobile GPU and its first Mali design with dedicated neural accelerators.
That design lets AI processing happen closer to the graphics workload instead of moving data to a separate AI processor. Arm claims the GPU delivers up to four times the neural-graphics performance per watt and up to 14% higher conventional gaming performance, based on its own testing.
The GPU supports Neural Super Sampling, which can reconstruct higher-resolution images from lower-resolution renders, and Neural Frame Rate Upscaling, which generates intermediate frames. Arm says Neural Super Sampling can reconstruct a 1080p image from graphics rendered at 540p, while Neural Frame Rate Upscaling can generate intermediate frames to turn 30 fps output into a displayed 60 fps experience.
The approach could let phone makers deliver better visuals without simply increasing the amount of raw graphics processing. Rendering fewer pixels and reconstructing the final image could give device makers more room to balance visual quality, power consumption and thermal limits.
But the technology comes with an important caveat: generated frames and reconstructed pixels still need to look convincing and respond quickly. Higher frame-rate numbers alone do not guarantee a better gaming experience.
The CPU becomes the AI traffic controller
Arm is also targeting the less visible work behind AI agents. The C2 CPU cluster combines C2-Ultra and C2-Pro cores and doubles the platform’s SME2 processing capability for on-device AI. Arm says C2-Ultra delivers up to 1.7 times the AI performance of C1-Ultra, 15% higher single-thread performance and up to 38% less power at the same performance.
The company says the additional CPU and SME2 capacity is designed for agentic workloads that coordinate context, applications and multiple on-device tasks. They claim the expanded SME2 capability can deliver up to a 70% speedup on selected small language models.
The bigger bet is coordination
What makes CSS for Mobile 2 interesting is that Arm is not treating AI as a single accelerator problem.
The company is combining AI-aware CPU processing with neural graphics and software tools, including its Neural Graphics Development Kit and AI Portal. That could make the platform more useful to developers who need AI features to interact with applications, graphics and other device functions rather than run in isolation.
Still, Arm’s licensees must incorporate the platform into commercial silicon before smartphone makers can ship devices. Until partners announce chips and products, their timing, cost and real-world performance remain open questions.
Read more: Samsung and Arm are working on 2nm chips designed to expand AI processing directly on devices.