Arm Puts Neural Accelerators Inside Its New Mobile GPU Shader Cores

Neural acceleration moves into the shader core

Arm announced CSS for Mobile 2 on Monday: a compute subsystem pairing the new C2-Ultra CPU with the Mali G2-Ultra NX, which the company calls its first AI-native Mali GPU. The architectural change is the interesting part. Mobile AI has traditionally lived in a separate NPU block that exchanges data with the GPU and CPU over shared memory. The G2-Ultra NX instead integrates neural acceleration directly into the graphics pipeline, with matrix hardware inside the shader cores that shares GPU memory and control structures with graphics work. [1] [2]

Arm says the design followed five years of development, and it is debuting alongside Xiaomi's XRING O3 chip as the showcase silicon. The GPU ships in configurations up to 16 cores with Vulkan 1.4 support and 4MB of L2 cache, and it introduces Arm's Neural Super Sampling, an AI upscaling technique for rendered frames. [3] [4]

The numbers Arm is claiming

The headline figures: up to 24 percent higher benchmark performance and 14 percent higher non-AI gaming performance from the GPU's architectural changes, up to 1.7 times higher on-device AI performance from the platform alongside the C2-Ultra CPU, and up to four times the performance per watt on neural graphics workloads. Arm also claims its third-generation ray tracing unit cuts ray tracing workload costs by up to 70 percent, and the C2-Ultra CPU brings SME2, the Scalable Matrix Extension, Arm's most significant instruction set addition in years. [1] [5] [6]

Why in-shader AI matters for on-device models

The reason this matters for local AI is data movement. On current phones, a frame or a tensor often travels between the GPU, the NPU, and shared memory, and every hop costs time and power. Running neural work inside the shader cores means graphics and AI share the same execution resources and memory pool, which is exactly the profile that on-device models need: generate a frame, run a model, composite the result, without copying buffers between accelerators. [1]

For the kind of workloads this site covers, on-device transcription, image understanding, and small language models, the practical effect should show up as headroom. A phone that spends less power per AI operation can run bigger models, or the same models for longer, on the same battery. Arm built CSS for Mobile 2 explicitly for agentic AI workloads, which is a signal about where the company expects phone makers to compete next. [2]

What to watch before phones ship

The platform is expected to reach Android phones from 2027, so none of this is testable today, and Arm's numbers are vendor claims measured on Arm's own configurations. The details worth tracking are which chipmakers license the full subsystem, how the in-shader accelerators are exposed to developers through Vulkan and neural APIs, and whether real-world models see the promised efficiency or only synthetic benchmarks do. Xiaomi's early adoption suggests Chinese vendors will ship first. [4] [6]