Arm announced a pair of Compute Subsystem (CSS) platforms for different ends of the compute spectrum. The first is Arm CSS for Mobile 2 with a new CPU and GPU. The second is the latest Neoverse CSS N4. Now for some details.

Arm CSS for Mobile 2 Architecture

Arm’s latest mobile platform (Fig. 1) is optimized for artificial-intelligence (AI) applications with AI CPU and GPU subsystems. It targets smartphones and tablets but works for edge devices that need high performance, especially AI acceleration and game-level, cinematic graphics.

The compute clusters can be scaled using multiple cores to provide the desired performance and efficiency. The push for higher performance across the board is the desire to support agentic AI on the device rather than depending on the cloud.

GPU Uses an AI-First Architecture

The Mali G2-Ultra NX was designed for AI support from the ground up (Fig. 2). It’s the more important change to the platform because of how it utilizes AI.

Of course, GPUs are being used for AI acceleration in general. Improving the AI performance means that a GPU may be able to handle AI chores as effectively as a dedicated neural processing unit (NPU), thereby eliminating the need for the NPU. The neural and graphics processing pipelines have been integrated to let the GPU handle both.

The other pieces to the puzzle include the next-gen Ray Tracing Unit (RTU) and Arm’s Neural Dawn technology (Fig. 3). This adds AI support in a number of areas; e.g., Neural Super Sampling (NSS) for better image quality, Neural Frame Rate Upscaling (NFRU) to increase frame rates for high-resolution displays, and Neural Super Sampling and Denoising (NSSD) in the RTU to support advanced lighting features.

Leading game engines like Tencent Magic Dawn and Unity China’s Tuanjie Engine are taking advantage of the Mali G2-Ulra NX’s AI capabilities. Features like the Mali G2-Ultra NX’s NSSD support allows game developers to deliver a more realistic image based on a lower-resolution input (Fig. 4).

Where Does the Arm C2-Ultra CPU Fit?

The Arm C2-Ultra CPU pushes the envelope for mobile processor performance with added AI support (Fig. 5). Each core has two scalable matrix extension SME2 units that can provide up to 70% greater acceleration for small language models (SMLs) hosted on a device. The core delivers 1.7X higher AI performance while using 38% less power.

The architecture and implementation offer a 24% improvement over the C1 Ultra (Fig. 6).

Arm Neoverse CSS N4 Optimized for Efficiency

Arm Neoverse CSS N4 targets applications like storage servers, data processing units (DPU), and cloud servers, where power efficiency is important (Fig. 7). The Neoverse CSS V series is the complementary platform designed for maximum performance.

Designers can pack in as many as 128 cores/die. It should double the performance compared to the previous generation while providing 1.25X the performance per watt. Memory bandwidth has been bumped up by a factor of 1.75.