Arm's C2-Ultra CPU focuses on AI acceleration for 2027 Android phones

Arm says C2-Ultra delivers up to 15 percent higher peak performance than the C1-Ultra. An 8 percent gain comes from raising clock speed to 4.45 gigahertz, while the remaining 7 percent is from architectural improvements. At the same performance level, the new core uses 38 percent less power, partly because of the move to 2-nanometer manufacturing. It also supports up to 3 megabytes of L2 cache.

The pipeline width is unchanged from the C1-Ultra. Instead, Arm widened the execution window by 30 percent, allowing about 2,600 instructions to be in flight at once compared with 2,000 previously. The core has better out-of-order execution with smarter speculation and faster recovery, which lets it detect and execute code dependencies earlier by analyzing workload and data types. The goal is to reduce time waiting for data and move less data between caches, which also cuts power.

The other new cores, C2-Premium, C2-Pro, and C2-Nano, did not receive such sweeping updates, according to Android Authority. They are essentially last year's C1 cores tweaked for area, efficiency, and power on next-generation 2nm lines.

For AI, Arm claims the C2 configuration can provide up to a 1.7 times performance improvement in AI models by accelerating low-precision math and using lookup table instructions, or LUTi, to reduce memory bandwidth demands. The CPU cluster gains two Scalable Matrix Extensions, or SME2, units compared with one last year. Android Authority notes that Xiaomi's XRING O3 already supports dual SME2 units with last year's C1 architecture, so this is not a strict rule. SME2 code runs as part of the regular instruction pipeline, meaning developers do not have to do additional work beyond compiling the instructions. The two SME2 units execute in parallel rather than doubling throughput, and Arm's reference platform supports higher SME2 clocks this year, up to 3 gigahertz from 2 gigahertz.

The improved AI performance translates into real-world gains, according to Arm's benchmarks: speech-to-text latency falls by 40 percent and memory retrieval search latency drops by 41 percent, allowing the phone to recall user preferences and past interactions more quickly.