Arm CSS for Mobile 2: C2 Ultra and AI-Native G2-Ultra NX

Arm CSS for Mobile 2: C2 Ultra and AI-Native G2-Ultra NX

The shift from Generative AI to Agentic AI is changing what mobile computing hardware needs to optimize for. Instead of processing a single prompt and returning a response, an on-device agent must continuously perceive intent, retrieve context, reason, and execute actions across applications.

That workload is fundamentally different from conventional mobile AI inference. It is latency-sensitive, bursty, memory-intensive, and heterogeneous, requiring the CPU, GPU, specialized accelerators, memory subsystem, and software stack to operate as a coordinated system.

Arm’s response is CSS for Mobile 2, the company’s second-generation mobile computing subsystem, unveiled at its Everywhere China event. Arm positions the platform as an AI-native mobile computing architecture built around the new C2 CPU cluster, the Mali G2-Ultra NX, the SI L2 system interconnect, and an increasingly integrated AI software ecosystem.

Rather than treating CPU and GPU performance as isolated metrics, CSS for Mobile 2 focuses on system-level optimization. The objective is to reduce latency between heterogeneous compute resources while improving sustained performance and energy efficiency for two emerging workloads: Agentic AI and AI-native graphics.

🚀 C2 Ultra: A CPU Designed for Agentic AI

#

At the center of CSS for Mobile 2 is the C2 CPU cluster. The flagship reference configuration uses a 2+6 topology, combining two C2-Ultra cores with six C2-Pro cores. The cores share an L3 cache through the DynamIQ Shared Unit (DSU).

The C2-Ultra is Arm’s highest-performance mobile CPU core in this generation. Arm reports a 15% improvement in GeekBench 6.3 single-thread performance, a 12% increase in multi-thread performance, a 15% improvement in Speedometer 3.1, and a 12% reduction in application launch time.

At equivalent performance, the C2-Ultra reportedly consumes 38% less power than the previous-generation C1-Ultra.

These gains are not the result of frequency scaling alone. Arm’s architectural improvements focus on execution width, data movement, and branch prediction.

Larger execution engine

#

The C2-Ultra operates at frequencies above 4.45 GHz and delivers more than 15% higher peak performance.

More importantly, the core can keep more instructions in flight, increasing instruction-level parallelism (ILP). Improved speculation allows useful work to continue while dependencies are being resolved, reducing idle cycles and execution stalls.

For Agentic AI, this matters because the CPU is responsible for a large portion of the orchestration layer. It must coordinate model invocation, memory retrieval, application execution, and communication between heterogeneous accelerators without becoming the serial bottleneck.

More efficient data movement

#

The C2-Ultra also expands its ability to manage outstanding memory operations.

Larger Load/Store buffers allow the processor to track more concurrent memory requests, while improved prefetching attempts to bring predictable data into the cache hierarchy before it is explicitly requested.

The combined effect is reduced backend stall time and faster execution of workloads with irregular or memory-dependent access patterns.

This is particularly relevant to agent workloads, where small models, vector retrieval, application state, and orchestration logic can repeatedly move data through the memory hierarchy.

More accurate branch prediction

#

Branch prediction has also been improved across a range of workloads, including browser execution and application launch.

Stronger instruction fetching, target-address prediction, and faster recovery from mispredictions help maintain a more consistent instruction stream while reducing the resources wasted on speculative execution along incorrect paths.

For latency-sensitive workloads, these improvements matter because agent pipelines often consist of many short processing stages rather than one long compute-intensive kernel.

🧠 SME2 and LUTi Accelerate On-Device AI

#

The C2-Ultra’s AI capabilities extend beyond general-purpose CPU improvements through support for SME2 (Scalable Matrix Extension 2).

Agentic AI does not necessarily require a single large model to perform every task. A practical on-device agent can instead combine multiple specialized models for speech recognition, memory retrieval, classification, reasoning, and other functions.

These smaller models are precisely where CPU-side matrix acceleration can become valuable.

SME2 increases computational density for suitable matrix-oriented workloads, while the C2-Ultra adds LUTi (Lookup Table) instructions for lookup-intensive operations.

LUTi is particularly relevant to extremely low-precision quantized models, where reducing repeated memory accesses can alleviate pressure on memory bandwidth.

Complementary optimization for encoding and decoding

#

The two capabilities target different parts of the inference pipeline.

During encoding, computation can dominate execution, making SME2’s higher compute density valuable. During decoding, memory bandwidth often becomes the limiting factor, making efficient lookup and data movement increasingly important.

This creates a complementary optimization path:

• SME2 improves compute throughput for matrix-heavy operations.
• LUTi reduces memory-access overhead in lookup-intensive workloads.
• Together, they address both compute-bound and bandwidth-bound portions of small-model inference.

Arm reports that a configuration using two C2-Ultra cores with SME2 achieves an average 1.7× performance improvement over two C1-Ultra cores with SME2 across speech recognition, personal memory retrieval, and reasoning workloads.

The tested workloads include Moonshine and Parakeet for speech recognition, LFM 2.5-ColBERT-350M for multilingual memory retrieval, and several small reasoning models, including LFM 2.5 1.2B, Qwen 3.5 0.8B/2B, MiniCPM 5 1B, and Gemma 4 E2B.

Agentic AI latency

#

Arm also demonstrated the C2-Ultra using an agent workflow based on planning an anniversary dinner.

Compared with the C1-Ultra:

• Speech-to-text was 40% faster.
• Memory retrieval was 41% faster.
• Structured prompt reasoning was 25% faster.

Overall processing time fell from approximately 450 ms to 280 ms, while enabling SME2 reduced latency by an additional 24% in the demonstrated workflow.

The significance is less about a single benchmark number than the cumulative effect of reducing latency across a serial pipeline. When perception, retrieval, reasoning, and execution are chained together, improvements at each stage compound into a more responsive user experience.

Flexible CPU configurations

#

Arm is also providing configuration flexibility across the C2 family.

Partners can choose configurations ranging from entry-level 2N designs to flagship 2U+6Pro implementations. The Ultra, Pro, and Nano microarchitectural tiers can be mixed within a cluster, with different core classes operating at independently scalable frequencies.

This allows SoC designers to balance performance, area, power, and product positioning rather than adopting a single fixed CPU configuration.

🎮 Mali G2-Ultra NX: Arm’s First AI-Native Mali GPU

#

CPU-side AI acceleration addresses agent orchestration, but mobile graphics face a different challenge.

Modern games increasingly combine high-resolution rendering, large virtualized environments, complex shaders, and ray tracing. Yet mobile GPUs operate under strict thermal and power constraints, often in the range of only a few watts.

Rendering every pixel natively at every frame becomes increasingly difficult under those constraints.

Arm’s answer is Neural Graphics: render only the most important information natively and use neural reconstruction to generate additional image detail, frames, and ray-tracing information.

The Mali G2-Ultra NX is Arm’s first Mali GPU designed around this AI-native approach.

Its architecture combines neural acceleration, a redesigned execution engine, and a third-generation ray-tracing unit.

Neural Accelerator integrated into shader cores

#

The Neural Accelerator (NX) is integrated directly into the GPU’s shader architecture rather than operating as a completely separate accelerator.

The unit supports INT8 and INT16 computation, tensor operations, weight compression, and operates at up to twice the GPU frequency. It is accompanied by a motion engine and optical-flow acceleration.

This tight integration allows neural graphics workloads to share GPU cache and memory infrastructure with traditional graphics and compute workloads.

The architectural benefit is reduced data movement between independent processing blocks. Instead of repeatedly transferring intermediate results between separate accelerators, neural operations can remain much closer to the graphics pipeline.

New execution engine

#

The Mali G2-Ultra NX also introduces a substantially redesigned Execution Engine (EE), representing the largest ISA-level change since Arm’s seventh-generation GPU architecture.

The design specifically targets workloads associated with Unreal Engine 5 technologies such as Nanite virtualized geometry and Lumen dynamic global illumination.

One important problem is register spilling. As shaders become larger and more complex, insufficient register capacity can force temporary values into memory, significantly reducing performance.

The new execution engine doubles register capacity per Warp and supports dynamic wide-register allocation at a 16-Warp granularity.

The goal is to keep more complex shaders resident in efficient execution resources while reducing spilling and preserving GPU headroom for larger scenes and more sophisticated visual effects.

Third-generation ray tracing

#

The third-generation Ray Tracing Unit (RTU v3) addresses another major mobile graphics bottleneck: scene geometry.

As triangle counts increase, conventional BVH structures can consume substantial cache and memory capacity. RTU v3 introduces a more compact triangle representation that avoids redundant storage of shared vertices and duplicate edges.

This increases the amount of useful geometry that can fit into cache while reducing DRAM bandwidth requirements.

The architecture also introduces hardware-level Opacity Micromaps (OMM), improving ray-hit testing for transparent and semi-transparent geometry such as foliage and fences.

Arm reports that OMM can increase frame rates by up to 30% while reducing the ray-tracing workload for applicable scenes by up to 70%.

✨ Neural Graphics: NSS, NFRU, and NSSD

#

The G2-Ultra NX uses its neural architecture to introduce three primary neural graphics technologies: Neural Super Sampling (NSS), Neural Frame Rate Upscaling (NFRU), and Neural Super Sampling and Denoising (NSSD).

These techniques target different stages of the graphics pipeline.

Neural Super Sampling

#

NSS renders at a lower base resolution and uses a neural network to reconstruct a higher-resolution image.

Arm demonstrates reconstruction from a 540p input to a 1080p output, reducing the number of natively rendered pixels to approximately one-quarter of native 1080p rendering.

Under sustained power constraints, Arm reports a 1.9–2.1× frame-rate improvement, a 50% reduction in DRAM bandwidth overhead, and a 50%–60% reduction in dynamic energy per frame.

The key advantage is that the GPU spends expensive native rendering resources only where they provide the most visual value, while the neural pipeline reconstructs additional detail.

Neural Frame Rate Upscaling

#

NFRU uses color, depth, and motion-vector information from consecutive native frames together with hardware optical-flow acceleration.

The neural model generates intermediate frames between native frames, allowing the perceived output frame rate to increase without requiring the GPU to render every frame fully from scratch.

In Arm’s testing, NFRU increased sustained game output from 60 FPS to 120 FPS while reducing DRAM access traffic by approximately 33%.

This approach is particularly useful when rendering throughput, rather than image reconstruction quality, is the dominant limitation.

Neural Super Sampling and Denoising

#

NSSD targets ray-traced workloads.

Mobile ray tracing often operates at relatively low sample counts because of power and compute constraints. The resulting images can contain substantial high-frequency noise.

Instead of treating super sampling and denoising as separate operations, NSSD combines them into a neural reconstruction stage that simultaneously increases image quality and removes ray-tracing noise.

This allows the platform to produce high-resolution ray-traced output without requiring the GPU to calculate every sample natively.

🕹️ Neural Graphics in Real Games

#

Arm has open-sourced its NSS and NFRU models through Hugging Face and GitHub and provides integrations for engines such as Unreal Engine.

Arm estimates that adapting existing games can take approximately 3–6 months, while new game development targeting the technology can require less than 18 months.

The company also demonstrated Neural Dawn, a mobile title co-developed with Sumo Digital.

The game uses Unreal Engine 5.5 MegaLights and supports as many as 1,400 dynamic light sources without relying on pre-baked lighting.

On the previous-generation CSS G1-Ultra platform, the workload reportedly ran at approximately 15 FPS. On Mali G2-Ultra NX, Arm demonstrated sustained performance of 60 FPS.

That represents a claimed 4× frame-rate increase, 70% higher energy efficiency, and 70% lower DRAM traffic.

Arm says Neural Dawn is scheduled to launch in 2026 alongside the first devices based on Mali G2-Ultra NX.

Traditional graphics performance

#

The G2-Ultra NX is not limited to neural rendering improvements.

Arm reports a 9% improvement in both traditional rasterization and baseline ray-traced game performance. Across selected benchmarks, the reported gains include:

• Steel Nomad Light: +17%
• InVitro 2: +18%
• AnTuTu v11: +20%
• Solar Bay Extreme: +24%

Overall, Arm reports approximately 20% higher average gaming performance, while single-frame DRAM bandwidth requirements in ray-tracing workloads fall by 14%.

🔗 SI L2: Connecting the Compute Subsystem

#

Increasing CPU and GPU performance is only useful if the rest of the SoC can move data efficiently.

CSS for Mobile 2 therefore replaces the previous SI L1 interconnect with SI L2, a channelized system interconnect designed to connect the C2 CPU cluster, Mali G2-Ultra NX, unified memory, and peripheral devices.

This layer is particularly important for AI workloads because the CPU and GPU frequently need to share intermediate data.

Bandwidth and memory

#

SI L2 supports LPDDR6, providing the bandwidth required by increasingly compute-intensive mobile workloads.

Higher memory throughput is important for both AI inference and graphics, but bandwidth alone does not determine system responsiveness.

Memory latency

#

Arm reports approximately a 45% reduction in demand-load latency between the CPU and DRAM compared with SI L1.

This is particularly significant for Agentic AI, where multiple short operations can be serialized by memory dependencies.

Reducing the latency of individual memory transactions can therefore have a greater effect on perceived responsiveness than simply increasing peak compute throughput.

Coherency and QoS

#

SI L2 provides scalable, hardware-enforced coherency and advanced quality-of-service mechanisms for different traffic classes.

This allows CPU and GPU workloads to share memory while maintaining predictable access behavior under concurrent load.

For heterogeneous AI workloads, this coordination is critical because independent accelerators may otherwise compete for the same memory and interconnect resources.

Security and physical efficiency

#

The interconnect also supports technologies such as Memory Tagging Extension (MTE) and Distributed Virtual Memory (DVM) while targeting low area and power overhead.

The resulting architecture is intended to balance performance, security, coherency, and physical implementation constraints within a mobile SoC.

🛠️ Software Ecosystem for Out-of-the-Box AI

#

Hardware acceleration is only useful when software can expose it efficiently.

Arm has therefore built a software stack around the C2 and G2-Ultra NX rather than treating the hardware as an isolated IP block.

KleidiAI

#

Arm’s KleidiAI serves as a core hardware-software optimization layer for AI workloads.

Arm reports that it has been integrated into more than 100 third-party applications and adapted to more than 12 major AI frameworks.

The stack includes more than 200 optimized microkernels and approximately 140,000 lines of code, providing optimized execution paths for Arm CPUs.

Models such as Qwen3-TTS and HY-MT1.5-1.8B-2Bit can benefit from these optimizations and execute on the C2-Ultra’s SME2 resources.

Arm AI Portal

#

The new Arm AI Portal provides a centralized environment for developers working with on-device AI.

It includes optimized models, performance and accuracy measurements, sample code, and deployment resources. MCP server integration also targets Agentic AI development workflows.

Developers can evaluate models against Arm hardware, compare workload characteristics, and move more quickly from model selection to deployment.

Arm is also providing early-access optimization tools for developers working with proprietary models.

Arm Neural Graphics SDK

#

For graphics developers, the Arm Neural Graphics SDK (ArmNG SDK) provides open models for NSS and NFRU alongside Vulkan ML extensions, Unreal Engine plugins, and custom-engine SDKs.

The toolkit also includes profiling, graphics optimization, training, and model-tuning capabilities.

This software layer is essential because neural graphics is not simply a hardware feature. Game engines, rendering pipelines, model execution, and developer tooling all need to expose the underlying acceleration efficiently.

🌐 From Faster Components to System-Level AI

#

CSS for Mobile 2 reflects a broader change in Arm’s mobile computing strategy.

Agentic AI and AI-native graphics have different performance requirements, but both expose the limitations of optimizing individual IP blocks independently.

For Agentic AI, the critical metric is end-to-end latency. Speech recognition, memory retrieval, reasoning, and application execution form a serial pipeline in which delays accumulate. Arm therefore emphasizes C2-Ultra CPU performance, SME2, LUTi, and SI L2 memory latency.

For AI-native graphics, throughput and energy efficiency dominate. Rendering every pixel and every frame natively is increasingly expensive under mobile power limits. The G2-Ultra NX instead combines selective rendering with neural reconstruction, motion processing, and ray-tracing acceleration.

The common denominator is system-level coordination.

CPU performance, GPU throughput, memory bandwidth, memory latency, cache utilization, interconnect coherency, accelerator integration, and software optimization all contribute to the final user-visible result.

This is the central architectural message behind CSS for Mobile 2: in the AI-native mobile era, simply combining a faster CPU with a faster GPU is no longer sufficient. The performance ceiling increasingly depends on how efficiently the entire compute subsystem moves data, schedules heterogeneous workloads, and keeps specialized resources fed.

The C2-Ultra and Mali G2-Ultra NX therefore represent more than two new processor blocks. Together with SI L2 and Arm’s software stack, they form a platform-level attempt to make mobile devices capable of running increasingly complex AI agents and neural graphics workloads within strict power and thermal constraints.

The ultimate test, however, will come from commercial flagship SoCs. Real-world benchmarks will determine how effectively this tightly integrated architecture translates Arm’s claimed improvements into sustained mobile performance, responsiveness, and energy efficiency.