Arm's AI-native chips claim 24% faster on-device agents, neural GPU
The Sept 8 launch spans phones and data centers: a first AI-native mobile GPU with a free neural-graphics kit, plus a 128-core cloud chip for agentic work.
Copy markdown
C2-Ultra: up to 1.7x AI, 24% faster agentic work
Arm's new flagship CPU core, the C2-Ultra, claims up to 1.7x higher AI performance and finishes agentic operations about 24% faster than last year's C1-Ultra, at up to 38% lower power. It's IP for 2027 flagship SoCs, so the gain reaches you through your next phone rather than a download — but it's a clear marker for where on-device agent latency is headed.
A first ‘AI-native’ mobile GPU — with a free dev kit
The Mali G2-Ultra NX bakes neural accelerators into the GPU shader cores: up to 4x performance-per-watt for neural graphics and roughly 70% less external memory traffic. Arm shipped the free Arm Neural Graphics Development Kit alongside it — neural super-sampling (NSS), neural frame generation, Vulkan ML extensions, and Unreal Engine plugins. Graphics and game devs can target neural upscaling now; Where Winds Meet and Infinity Nikki are among the early adopters.
Neoverse CSS N4 ‘Ranger’: the cloud your agents run on
The new server platform scales 8–128 cores per die at up to 3.8GHz for about 2x the socket performance of N3, with PCIe Gen6/7 and CXL 4.0 — pitched squarely at agentic-AI and inference in the data center. It's the blueprint for the next wave of cheap Arm cloud instances that a lot of your inference already runs on.
The catch: it's IP, and the matrix units skip FP8/BF16
None of this is downloadable — Arm licenses these designs to chipmakers, so real silicon lands in 2027 devices, not this quarter. And the GPU's matrix accelerator does INT8/INT16 only (no FP8 or BF16), so the on-device wins tilt toward neural graphics rather than large-LLM inference. Temper the ‘run big models on your phone’ expectations accordingly.
Elsewhere: Arm's Physical-AI stack, and Mistral's €3B
Same day, Arm launched Total Design for Physical AI (80+ partners including AWS and NXP) with a Robotics Capability Framework that grades autonomy from RL0 to RL5 — scaffolding for humanoids and robots. Separately, Mistral closed a roughly €3B round near a €21B valuation, Samsung-led; it's money rather than a shippable release, but it keeps an open-weight frontier lab well funded.