Arm's CSS for Mobile 2 Puts AI into CPU and GPU, Leaves NPU to Partners

The CPU announcement centers on SME2, Arm's matrix-accelerator instruction set for AI. Arm said SME2 has become standard in virtually all flagship smartphones, both Android and iOS. The new C2 Ultra provides 15 percent higher single-thread and webpage-browsing performance and 12 percent faster app-launch and multi-thread performance. In selected AI models, the SME2-enhanced C2 CPU cluster executes up to 1.7 times faster than the previous generation. Arm positions the CPU as the execution point for lightweight models, tool calling and agent orchestration, where per-step latency affects the overall experience. Daniel Lu, Arm's director of edge AI CPU product management, told Leiphone that CPU clusters with SME offer equivalent performance of about 5 to 6 TOPS, and that the company prioritizes low latency over peak throughput. The C2 cluster adds private L2 and large shared L3 caches to cut DRAM accesses and data-movement power.

With the Mali G2-Ultra NX, Arm integrates a neural accelerator into a Mali GPU for the first time. The GPU can use neural super-sampling to render at lower resolution and reconstruct high-resolution frames, neural super-sampling plus denoising for ray-traced scenes, and neural frame-rate up-conversion to insert AI-generated frames between real frames. In a live demo, Arm showed up to 4 times the frame rate and energy efficiency versus the previous Mali G1-Ultra rendered without AI assistance, and DRAM traffic reduced by up to 70 percent. Without neural graphics, Arm lists an average 20 percent gain in synthetic game benchmark performance and up to 14 percent higher frame rates in real-world same-frequency gaming tests, illustrating the headroom created by AI reconstruction. A new execution engine and third-generation ray tracing unit use a more compact triangle data format to reduce duplicate data, cutting DRAM bandwidth needs by about 13 percent in existing ray-tracing benchmarks. The GPU can also run non-graphics neural networks in supported data formats such as INT8, but Arm's immediate focus remains graphics.

Absent from CSS for Mobile 2 is an Arm-designed NPU. James McNiven, Arm's vice president for edge AI and intelligent computing terminals, said at the event that in mobile devices Arm is accustomed to letting partners lead NPU-level innovation, while its Ethos NPU line is used in IoT. Many phone-chip partners have proprietary NPU architectures and optimize them together with their models, operating systems and first-party AI features. For third-party developers, CPU and GPU provide more standardized execution environments across brands, and system software decides whether a workload runs on CPU, GPU or NPU, Arm said. The company said CSS for Mobile 2 offers configurable compute: the C2 cluster supports up to 14 cores, allowing customers to mix C2 Ultra and C2 Pro cores according to workload, and the Mali G2-Ultra NX can be configured with different numbers of shader cores and NX units.

Arm's software stack is now a prominent part of the platform. KleidiAI connects SME2-optimized kernels to AI frameworks, letting developers exploit the hardware without writing low-level instructions. For GPU neural graphics, Arm has opened its neural graphics models and offers SDKs, sample code and game-engine plugins, with Tencent Games and Unity China listed as partners in adaptation work. Arm's executives said the company aims to shorten the distance between processor capabilities and developer applications and to make CSS a system-level offering rather than a list of IP specifications.