Click the blue text above 【AnTuTu】 to follow meSharing technology hotspots every day!
Recently, Arm officially launched the new Arm Lumex Compute Subsystem (Lumex CSS) aimed at mobile devices at its UNLOCK summit in Shanghai. This platform is based on the Arm V9.3 instruction set and includes the C1 series CPU clusters (C1-Ultra, C1-Premium, C1-Pro, C1-Nano) as well as the next-generation Mali G1 GPU family (Mali G1-Ultra, Mali G1-Premium, Mali G1-Pro).
One of the cores of Arm Lumex is the upgrade of the matrix extension in Armv9.3 — SME2 (Scalable Matrix Extension 2).
Arm emphasizes the improvements in throughput and power efficiency brought by SME2 in matrix and vector operations, supporting dynamic expansion and dequantization of 2-bit and 4-bit quantized weights, thus achieving more efficient AI inference on memory bandwidth-constrained mobile devices.
Arm stated in its release materials that the C1 series with SME2 enabled can achieve multiple times the AI performance improvement and several times the energy efficiency improvement under various typical workloads such as generative AI, speech recognition, and computer vision compared to the previous generation.
In terms of CPU segmentation, C1-Ultra is positioned as a super core aimed at flagship and high single-thread performance, with Arm providing target values including significant IPC improvements and front-end/prefetch optimizations for real workloads; C1-Premium aims for “optimal PPA (Performance/Power/Area)” as its design goal, suitable for replacing super cores in area-constrained but still high-performance sub-flagship/high-performance products for area and cost optimization; C1-Pro focuses on high-efficiency large cores, while C1-Nano targets wearable and ultra-low power scenarios.
Some benchmarks listed at the press conference show the average performance improvement and power reduction ratio of the new generation C1 cluster in common mobile scenarios (gaming, video, daily applications).
Arm claims that the overall performance of the new platform has improved by about 30% compared to the previous generation in several benchmarks, with an average speedup of about 15% in daily application scenarios (gaming, streaming), and a power reduction of about 12% under the same conditions. In AI inference, the C1 cluster with SME2 enabled can achieve “up to 5 times the AI performance improvement” and “up to 3 times the energy efficiency improvement” claims.
The Mali G1 series aims to bring “desktop-level ray tracing and edge AI” to mobile phones and lightweight terminals.
The G1 brings improvements in both image quality and inference through a new ray tracing unit (RTUv2), a matrix multiplication path for FP16 (MMUL/FP16), and a restructured rendering pipeline, while balancing energy efficiency and scalability. The core hardware design of Mali G1-Ultra provides dedicated acceleration paths for real-time ray tracing and AI inference.
The next-generation ray tracing unit RTUv2 supports more efficient BVH (Bounding Volume Hierarchy) traversal and ray solving in hardware. Arm claims that compared to the previous generation Immortalis-G925, RTUv2 can bring up to about 2× improvement in ray tracing performance and achieve higher average frame rates in hardware ray tracing games.
The implementation of RTUv2 is more focused on “specialized computation for single ray paths” to achieve more stable real-time performance in memory and cache-constrained scenarios on mobile devices.
In terms of AI capabilities, Mali G1 introduces an MMUL path for FP16 and a series of matrix computation optimizations, providing significant acceleration for common mobile inference tasks such as semantic segmentation, image denoising, depth estimation, object detection, and speech recognition. Arm claims that the inference speed of the GPU can improve by about 20% compared to the previous generation on certain networks and workloads, and in specific tasks (such as certain image/denoising scenarios), it can even reach close to or exceed 100% improvement. This means that manufacturers can offload more AI tasks to the GPU and RTU in SoC design, reducing reliance on independent NPUs.
Architectural scalability and system-level optimization are also one of the selling points of G1.
Arm claims that G1-Ultra can scale up to a 24-core configuration and has made improvements in L2 cache, interconnect scheduling, and pipeline restructuring to reduce memory bottlenecks when graphics and AI are running concurrently. Arm also emphasizes the improvements in power consumption/frame energy consumption of G1: compared to the previous generation, energy consumption per frame can be reduced by about 9%.
Everyone is watching:
Is the new generation of “nail households” here? What upgrades does the iPhone 17 actually have?
August Android performance rankings: Snapdragon 8 Gen 2 remains strong
August Android cost-performance rankings: Dimensity 9400+ only needs a little over 2000Quickly filter and order: AnTuTu’s device selection mini-program is officially launched
↓↓↓ How to test benchmark scores? Click to read the original text