Arm C2-Ultra Delivers 1.7x AI Performance on Mobile

Arm's New Mobile Platform, Component by Component

Arm announced CSS for Mobile 2 on September 8, and the interesting part is not a single core but the way the pieces are matched. The platform bundles two CPU designs — the C2-Ultra and the C2-Pro — with the Mali G2-Ultra NX GPU and a supporting software stack including KleidiAI and a Neural Graphics Development Kit. Both CPUs are built on Armv9.3-A with Armv9.4-A extensions.

• C2-Ultra: up to 1.7x higher AI performance and 15% higher single-thread performance than the C1-Ultra, at up to 38% less power for the same performance
• Doubled SME2 units: Arm claims a 70% speedup on recent small language models
• Mali G2-Ultra NX: the first Mali GPU with dedicated neural accelerators, configurable from 10 to 24 cores
• Graphics: up to 4x better performance per watt on neural graphics, and up to 14% higher performance on existing game content

Cluster configurations reach 14 CPU cores with up to 32MB of L3 cache, 64KB instruction and 128KB data caches per core. The GPU supports OpenGL ES 3.2, Vulkan 1.4 and OpenCL 3.0.

Why Neural Accelerators on the GPU Are the Real News

Putting matrix units inside the GPU rather than beside it is a specific architectural bet. The Mali G2-Ultra NX runs neural and traditional graphics work in the same pipeline, which is what makes techniques like Neural Super Sampling and Neural Frame Rate Upscaling practical on a phone power budget rather than a desktop one. Arm's supporting figure is a 70% reduction in DRAM traffic with neural frame rate upscaling — and on mobile, memory traffic is very often the actual thermal constraint, not arithmetic.

That matters beyond phones. The same IP shows up in the single-board computers and compact Arm systems we follow in our mini PC and SBC coverage. When an accelerator moves into the GPU block, it tends to reach cheap silicon a generation or two later, which is roughly the path that brought NPUs to boards like the Qualcomm Dragonwing Q-2390.

What Does 1.7x AI Performance Actually Buy You?

Worth being precise: the 1.7x figure is Arm's own generation-over-generation comparison against the C1-Ultra, not an independent benchmark, and the 70% small-language-model speedup comes from doubling SME2 capacity rather than from a new instruction set. Both are CPU-side numbers.

Practically, the doubled SME2 units are the line item most likely to change what runs locally. Small language models in the 1B-to-8B range are exactly the workload that has been sitting just past comfortable on mobile silicon — usable, but slow enough that developers route to a server instead. A 70% improvement does not settle that argument, but it moves the threshold, and it does so on the CPU, which every application can reach without a vendor-specific runtime.

When Do Devices Ship?

Arm has not published a general availability date for the platform itself, which is normal for IP licensing. Silicon timing is the useful signal: per CNX Software, flagship designs from MediaTek are expected first, with Samsung's Exynos 2700 following in early 2028. Arm's own announcement lists Tencent Games, Unity China, NetEase, Google, OPPO and vivo among ecosystem partners, and points to NetEase's Where Winds Meet as an early neural graphics deployment.

For anyone building around Arm edge hardware, the sensible read is to treat the Neural Graphics Development Kit as the near-term item and the silicon as a 2027 planning question. If you are sizing local AI workloads today, our guide to how much RAM a local LLM needs is the more immediately actionable piece.

Sources: Arm Newsroom — September 8, 2026; CNX Software — September 8, 2026.