Arm Launches Lumex Chip Series Optimized for Mobile AI

Arm Holdings plc has unveiled a new lineup of chip designs today, intended for smartphone manufacturers developing mobile processors.

This series, named Lumex, features central processing units, graphics processing units, and an interconnect module. Smartphone manufacturers can use this interconnect module to integrate Arm's new GPU and CPU into a single system-on-chip.

The Lumex lineup includes four CPU core designs utilizing the company's SME2 technology. This instruction set extension comprises a set of low-level computing operations, enabling chips to perform complex calculations by combining these operations. The computing operations in SME2 are optimized to efficiently run artificial intelligence models.

Mobile developers often compress AI models in applications to enhance performance and reduce battery consumption. According to Arm, SME2 includes a 512-bit register optimized specifically to support compressed neural networks. A register is a high-speed memory pool directly embedded within the CPU.

SME2 also introduces new computing operations for data movement and modification. Additionally, the so-called predicates act as counter mechanisms, reducing the amount of data that needs to be processed to complete computing operations. This reduction can help improve throughput in certain cases.

Within the Lumex chip design family, the most powerful of the four CPU core designs is called C1-Ultra. According to Arm, this processor offers a 25% improvement in single-threaded performance compared to the previous generation chip. This makes it well-suited for hardware-intensive AI applications, such as content generation.

C1-Ultra is introduced alongside C1-Premium, which is designed to support simpler AI applications like voice assistants. C1-Premium is 26% smaller than C1-Ultra.

The other two C1 cores offer a compromise in performance to improve efficiency. C1-Pro is aimed at applications prioritizing sustained performance, while C1-Nano features a compact design suitable for wearables.

Smartphone manufacturers can flexibly combine Arm's new CPU cores. For example, a company could design a smartphone processor containing one C1-Ultra and multiple C1-Pro cores. High-demand applications can run on the C1-Ultra, while other tasks are assigned to other cores to save energy.

"For generative AI, speech recognition, classic machine learning (ML), and computer vision (CV) workloads, Arm C1 CPU clusters enabled by SME2 can achieve up to 5x faster AI acceleration under the same conditions compared to the previous generation CPU clusters," wrote Stefan Rosinger, Senior Director of CPU Product Management at Arm, in a blog post. "With SME2, CPU cluster efficiency improves by up to 3 times."

The C1 cores are designed to work with Arm's new Mali G1 GPU series launched simultaneously. The most advanced GPU in this series, G1-Ultra, delivers 20% faster inference speeds than its predecessor. A new ray tracing module enables it to render lighting, shadows, and reflections at twice the speed.

G1-Ultra is equipped with 10 shaders, which function similarly to cores in a CPU. It is introduced alongside three other G1 GPUs containing up to nine shaders each. Additionally, Arm has released a new interconnect module, S1 L1, which can connect C1 and Mali G1 chips, reducing static latency by 75% compared to earlier hardware.

Arm states that processors based on the Lumex chip designs can be manufactured using a three-nanometer process. The company expects the first mobile devices featuring Lumex to hit the market later this year or in early 2026.