Arm today introduced CSS for Mobile 2, an AI-native compute platform built around a new C2 CPU cluster and the Mali G2-Ultra NX GPU. Official announcement is here. Companion technical notes: C2 cluster and Mali G2-Ultra NX.
This is a platform IP deal, not a product SKU. Arm is selling a packaged CPU + GPU + system-IP + software stack to SoC partners. Devices land on partner calendars, not Arm’s.
The package
- C2 cluster: C2-Ultra (peak core) + C2-Pro (efficiency cores), Armv9.3-A with SME2. Arm says each of those cores carries two SME2 units — double the prior-gen matrix capacity.
- Mali G2-Ultra NX: first Mali GPU with dedicated neural accelerators wired into the shader cores, plus a new execution engine and third-generation ray tracing unit.
- Software: KleidiAI, Neural Graphics Development Kit, Arm AI Portal. Arm is pitching this as ready for the existing 22-million-developer stack, not a clean-room SDK.
The numbers Arm is putting on the board
- C2-Ultra vs C1-Ultra: up to 1.7x AI performance, 15% single-thread, 12% cluster multi-thread; up to 38% less power at iso-performance.
- Doubled SME2: ~70% speedup on current small language models.
- Speech-relevant figure from Arm’s own C2 write-up: C2-Ultra + SME2 cuts latency ~40% vs C1-Ultra on Moonshine and Parakeet speech-to-text.
- Mali G2-Ultra NX: up to 4x perf/watt on neural graphics; ~14% on existing (non-neural) game content; up to 24% on graphics benches.
Those are vendor benches. Treat them like spring-training exit velo, not Opening Day ERA.
Why this is on the audio-compute ledger
Arm did not ship an audio ASIC today. The graphics partners named (Tencent, Unity China, Sumo Digital, NetEase, Infold) are game-side. The audio tell is the CPU side: SME2 is matrix throughput on the application cores, which is where on-device ASR, voice agents, stem-ish enhancement, and small generative-audio models actually run when you cannot burn an NPU budget or a cloud round-trip.
A 40% STT latency cut on Parakeet/Moonshine is the closest thing in this release to a live-performance or on-device DAW number. Agentic AI on-device is the product story Arm wants; speech in / action out is the workload that pays for it on phones.
Comparables
- Last meaningful Arm audio-adjacent drop was Stable Audio Open Small on Arm (2025) — model port, not a new ISA.
- Qualcomm’s parallel track is Hexagon NPU + Adreno Matrix Cores (Neural Fusion, Sept 2) plus today’s AWS inference-silicon pact. Different layer: Arm is licensing the phone; Qualcomm is also trying to own the rack.
- Apple stays on its own AMX / AFM-on-device path. This Arm stack is the MediaTek / Exynos / everyone-not-Snapdragon-or-A-series lane.
What it does not do
No cluster allocation for Suno/Udio/Lyria. No licensed music corpus. No DAW or plugin host change. No shipping phone named. Notebookcheck-style reporting has MediaTek first, Exynos later; that is not in Arm’s release and should not be treated as official.
Bottom line
Arm just raised the floor on what a mid-to-high mobile SoC can spend on local matrix math and neural graphics without a discrete NPU tax. For music-tech, the trade is inference locality: voice-in control surfaces, on-device mix assistants, and small audio models that used to bounce to the cloud. The graphics silicon is the headline; the SME2 doubling and the STT latency claim are the parts that move the audio-compute board.