Next-Gen ARM Cortex-X Powers 3nm AI Chip Designs

ARM has recently introduced its fifth-generation Cortex-X core for smartphones, accompanied by a cutting-edge compute sub-system (CSS) featuring the latest high-performance graphics unit.

The smartphone compute sub-system (CSS) integrates the ARM Cortex-X925, the successor to the X4, and the Immortalis G925 GPU core. These components are combined with refreshed ARM A-class cores, the A725 and A520, tailored for big.LITTLE architectures and ready for chip makers. The system also includes a system management wrapper, the Dynamiq DS-120, which efficiently manages the data flow between up to 14 cores.

The primary focus of the CSS is to enhance AI capabilities in Android smartphones using chips like the Mediatek Dimensity 9400, as well as in Windows and Linux PCs. These chips are designed for the 3nm process at TSMC and Samsung, and the 1.8nm 18A process at Intel Foundry Services (IFS).

ARM has also rolled out a V3, N3 CSS for custom chip and chiplet designs. The company has introduced library software that supports popular generative AI frameworks such as Llama 3. There are two versions of the Kleidi library available to support generative AI and computer vision, both of which are compatible with the ARMv9.2 instruction set used by the X925 CPU cores.

KleidiAI seamlessly integrates with renowned AI frameworks like PyTorch, Tensorflow, MediaPipe Microsoft Phi-3, and Meta Llama 3, ensuring backward and forward compatibility. On the other hand, KleidiCV accelerates CV pipelines essential for various camera applications. OpenCV, the world's largest CV library, has reported a significant performance boost of 75% for a range of image processing tasks thanks to KleidiCV integrations.

Samsung is set to take the ARM Cortex-X to the 2nm level. The Cortex-X95 offers a 36% performance increase by boosting the frequency from 3.6GHz to 3.8GHz, along with a higher instruction per cycle (IPC) rate and a larger private L2 cache. The Immortalis GPU delivers a 37% improvement in graphics performance, featuring between 10 and 24 processing units, up from 16, and supporting ray tracing capabilities.

The integrated system-level cache (SLC) enhances system power efficiency by reducing DRAM bandwidth and accesses. Additionally, the System Memory Management Unit (SMMU) ensures enhanced security through stage-2 translation to support virtualized security frameworks like the Android Virtualization Framework (AVF).

According to Chris Bergey, SVP and GM of the Client Line of Business at ARM, the latest Arm compute platform offers over 30% increased performance for demanding real-life Android workloads, 59% faster AI inference, and enhanced graphics performance, catering to a wide range of AI/ML and computer vision (CV) tasks.

MediaTek, a key customer for ARM's Cortex-X cores in smartphones, is gearing up to launch a 3nm chip for volume production in 2024. MediaTek's commitment to supporting the latest Armv9 Cortex-X925 CPU and Immortalis-G925 GPU Client solution in their upcoming flagship chipset, Dimensity 9400, underscores their dedication to advancing computing technology.

The CSS has garnered support from major foundries with leading-edge process technologies, including Intel Foundry Services and Samsung Electronics. These collaborations aim to deliver top-tier silicon technologies integrated with cutting-edge computing solutions to meet the evolving demands of customers.

As the industry moves towards 3nm AI chips and 6nm microcontrollers, TSMC's Dresden fab is poised to play a crucial role. The AI-optimized Arm CSS exemplifies the successful collaboration between Arm and TSMC, empowering designers to achieve unparalleled levels of performance and power efficiency in semiconductor innovation for AI applications.

For more information, visit www.arm.com.