Arm has released the Lumex CSS,supporting 3nm process technology at the ARM UNLOCKED summit held today. Arm officially launched the Arm Lumex CSS computing subsystem aimed at mobile devices. This platform integrates advanced components, supports 3nm process technology, and achieves significant performance improvements, aiming to deliver a transformative AI experience for flagship smartphones and next-generation PCs.
The platform integrates the Armv9.3 CPU with SME2 technology, demonstrating outstanding CPU performance. The C1 CPU cluster achieves significant breakthroughs, with the SME2 unit enhancing matrix computation efficiency by 5 times and the Gemma 3B model inference speed increasing by 4.7 times. The flagship C1-Ultra features an innovative microarchitecture, providing a 25% increase in single-thread performance compared to the previous generation, with a 15% speed gain in scenarios like video streaming. It offers a rich CPU hierarchy, with C1-Ultra providing peak performance, C1-Premium focusing on area efficiency, C1-Pro achieving sustained efficiency, and C1-Nano suitable for wearable devices, catering to different device needs.
On the GPU side, the Mali G1-Ultra is equally impressive. Its RTUv2 ray tracing unit employs an innovative modular design and idle power-off mechanism, saving 15% energy for games like “Genshin Impact.” The single-ray model architecture boosts rendering speed for “Honkai: Star Rail” by 19%, while the FP16 matrix computation path accelerates AI inference speed by 20%. The expanded L2 cache design allows the GPU to process graphics and AI workloads in parallel, with performance in semantic segmentation tasks soaring by 104%.
In practical applications, the effectiveness of SME2 technology is remarkable. The text-to-speech generation speed for the “Smart Yoga Coach” application has increased by 2.4 times; in collaboration with Alipay and vivo, the response time for large language model interactions has been reduced by 40%. Moreover, this technology has been adopted by numerous leading ecosystem partners, including Alibaba and Tencent.
Additionally, the Lumex platform offers flexible options for partners. They can directly adopt the platform and customized physical implementation solutions to shorten time-to-market; or configure register transfer level designs for personalization. At the same time, its complete software stack and tools assist developers in easily achieving AI acceleration and performance tuning.
Core Metrics of Arm’s New Lumex CSS Platform
1. CPU-C1-Ultra: 25% performance increase and single-thread improvement over Cortex-X925, 15% faster in video streaming, enhanced wide microarchitecture
2. CPU-C1-Pro: 12% performance increase over Cortex-A725, 12% energy savings, optimized Cortex A75
3. CPU-SME2: C1 cluster includes SME2, 5 times matrix efficiency improvement, 4.7 times lower latency for Gemma 3B inference
4. GPU-Mali G1-Ultra: First to support next-generation mobile ray tracing, 20% performance increase, 9% energy efficiency improvement
5. System Architecture: SI L1 offers superior bandwidth and latency, SLC power consumption reduced by 71%, supports MTE for security
6. Software Ecosystem: Adopted by Alibaba, Tencent, and others
– End –