Arm unveils its designs for the latest generation of ARM-based CPUs and GPUs. The Cortex brand name disappears and Mali also takes a back seat: from now on, prefer C1 and G1.
ARM specialist Arm launches new designs for mobile processors. Arm’s designs form the basis of many mobile chips and are integrated under license by companies including MediaTek. The announcements fall under the new Arm Lumex flag. That’s the name for designs tailored for mobile use. PC designs will fall under Arm Niva.
The CPU designs within Lumex were traditionally called Cortex, but Arm is now doing away with that. Cortex A9xx, Cortex-A7xx and Cortex-A5xx become ARM C1-Ultra, C1-Pro and C1-Nano. Together they fall under the C1 CPU cluster.
The CPUs are built on the Armv9.3 architecture. It goes without saying that Arm claims this architecture is focused on on-device AI. How relevant the AI capabilities are in practice remains to be seen. More important for the daily performance of mobile devices are the power and energy efficiency of the CPU cores.
Four New CPUs
The C1 cluster consists of four CPU types:
- C1-Ultra: The most powerful core in the series. According to Arm, this offers up to 25 percent better single-thread performance than the Cortex-X925, with improved IPC (instructions per cycle), and optimizations in memory management and branch prediction.
- C1-Premium: 35 percent smaller than C1-Ultra, but with comparable performance on benchmarks. Focused on efficiency within a smaller footprint. This chip is new and doesn’t directly replace a Cortex CPU.
- C1-Pro: This chip is optimized for better performance per watt. Arm claims up to sixteen percent improvement in gaming and up to twelve percent energy savings in browsing and video.
- C1-Nano: This is the most compact core, designed for wearables and energy-efficient devices. Arm states that the chip is up to 26 percent more efficient than the Cortex-A520, while maintaining performance.
AI Sauce
Arm points out that all C1 CPUs are equipped with Scalable Matrix Extension 2 (SME2): a built-in AI accelerator. SME2 accelerates matrix operations and allows AI models to run locally without additional programming work, provided there’s support from popular AI frameworks such as Google LiteRT or ONNX Runtime.
Thanks to SME2, partners see, according to Arm:
- Up to 4.7 times faster performance on speech models like Whisper.
- 2.8 times faster audio generation via models like Stable Audio.
- Better AI responsiveness in multimodal apps and camera inference.
The C1-DSU – the shared control center of the cluster – supports scalability tailored to different types of devices. For example, chip builders can deploy a combination of C1-Pro and C1-Nano for midrange smartphones with AI functionality.
Mali Takes a Back Seat
Parallel to the new CPU series, Arm launches the Mali G1 GPU family, with the G1-Ultra as the top model. Mali is thus retained, but Arm aligns the GPU naming more clearly with that of the CPU, with G1 and C1 as GPU and CPU partners of the same generation.
The new GPU builds on the fifth generation Arm GPU architecture and offers improvements in both gaming and AI performance.
Even more AI
The G1-Ultra contains the RTUv2, a renewed ray tracing unit that delivers up to 2x better ray tracing than the previous generation Immortalis-G925. In graphical benchmarks, Arm itself sees an average 20 percent performance gain, and for AI tasks up to twenty percent faster inference via new FP16 matrix paths.
The GPU also supports new AI applications, such as object recognition and image enhancement, with up to 104 percent faster processing for some workloads. Developers can utilize these capabilities through improved tooling and support for Arm’s own upscaling technology (Arm ASR), available in Unreal Engine 5, among others.
Besides the Ultra variant, there are also G1-Premium and G1-Pro, with scalable performance for other segments. Chip manufacturers can choose between 1 and 24 shader cores depending on the intended device type.
Manufacturers to Take Action
With the C1 CPUs and G1 GPUs, Arm aims to provide its customers with designs for mobile devices that process AI locally and run demanding games smoothly, without excessive energy consumption. The extent to which these local AI capabilities are also relevant alongside AI in the cloud will depend on the implementation by smartphone manufacturers. The new hardware will appear in the coming generations of smartphones and other mobile devices.