At the recently held Arm Everywhere China annual conference, Arm officially launched its second-generation mobile computing subsystem (CSS for Mobile 2), providing comprehensive support for agent AI and AI-native graphics technologies within a single computing platform. The system integrates the new Arm Mali G2-Ultra NX GPU, the Arm C2 CPU cluster featuring SME2, enhanced system IP, physical implementation solutions, and a developer-ready software ecosystem, all fully optimized for next-generation mobile devices.
Among them, the Arm Mali G2-Ultra NX is the first Mali GPU to integrate a dedicated neural network accelerator, tightly merging neural network computation with traditional graphics processing to achieve higher energy efficiency and responsiveness. This GPU can achieve up to a 4x improvement in "performance per watt" in neural network processing scenarios, and compared to the previous generation, it can deliver up to a 14% performance boost in existing game content. To demonstrate its performance, Arm also showcased on-site the game "Light and Shadow Reborn," jointly developed with Sumo Digital. It features cinematic lighting effects previously unattainable in mobile games, as well as smoother, higher-resolution gaming experiences. In the future, NetEase's "Where Winds Meet" plans to launch a version supporting neural super sampling for players this year, and Tencent's "Arena Breakout: Infinite" is also co-developing neural super sampling and denoising (NSSD) technology demonstrations with Arm, bringing revolutionary upgrades to the gaming experience.
In addition, Arm announced the new Arm C2 CPU cluster at this conference, combining Arm's most powerful mobile CPU—the C2-Ultra, the energy-efficiency-optimized C2-Pro CPU, and dual SME2 units. This enables AI workloads to be processed with high responsiveness and low latency directly on the CPU, while seamlessly collaborating with various AI accelerators. The C2 CPU cluster features doubled SME2 computing power, which can increase the running speed of the latest small language models by up to 70%, further enhancing the CPU's acceleration capability for AI workloads.
Finally, Arm also announced its second-generation mobile computing subsystem—CSS for Mobile 2. It is designed to address the trend where agent AI and neural graphics technologies are reshaping the capabilities required by mobile devices. Arm believes that the next generation of mobile AI experiences will no longer be defined by a single accelerator, but rather by integrating various AI capabilities within a unified, optimized system to create a new mobile experience.
After the conference, we were fortunate to join other domestic media outlets for an interview lasting over forty minutes with Chris Bergey, Executive Vice President of Edge AI at Arm, and James McNiven, Vice President of Edge AI Intelligent Terminal Computing at Arm. We discussed Arm's current product ecosystem and technical highlights, as well as their expectations for the entire industry and Arm's future development. Below is the detailed interview content.
Q: We understand that the Arm NX unit supports INT8 and INT16 precision. Some information indicates that the platform can achieve FP16 precision operations. Why is FP16 computation not handled by the NX unit? What are the considerations behind this?
James McNiven:The data formats supported by the Arm neural network accelerator itself are INT8 and INT16. This design aims to achieve higher efficiency while ensuring the precision required for application scenarios such as Neural Super Sampling (NSS), providing excellent energy efficiency and performance. At the same time, we also support floating-point format operations including FP16, but this capability is not located within the neural network accelerator; instead, it is handled by the standard shader cores in the execution engine. This allows for the best balance between specialized AI acceleration and general-purpose computing capabilities for different types of workloads.
Q: Regarding the neural network accelerator mentioned this morning, why is it placed inside the GPU rather than in the NPU?
James McNiven:We believe that different types of AI workloads require different forms of acceleration, and heterogeneous computing is the best way to meet diverse computational needs. Different application scenarios and computational intensities need to be handled by different types of compute engines. Therefore, the neural network accelerator integrated into our GPU is specifically designed to accelerate neural graphics workloads and is highly optimized for these tasks. Meanwhile, in our latest C2 CPU cluster, we have configured up to two SME2 units, which can accelerate AI workloads based on the CPU instruction set, providing faster response capabilities and lower latency. In short, we believe that each compute engine should have corresponding AI acceleration capabilities, allowing various workloads to run efficiently on the most suitable compute unit. Additionally, our chip partners typically integrate dedicated NPUs to accelerate their own specific AI applications and workloads.
Additionally, I want to emphasize one point: deploying SME2 within the CPU cluster provides extremely low latency and fast response capabilities, while also fully leveraging the existing security models and mechanisms of the CPU. Integrating the neural network accelerator inside the GPU allows it to share memory and cache resources of the execution engine, such as L1 and L2 caches, thereby processing AI graphics workloads more efficiently and achieving faster response times. The benefit of this design is that it not only delivers optimal performance and energy efficiency but also allows AI acceleration functions to closely collaborate with the graphics rendering pipeline, enabling unified scheduling and parallel processing with other workloads to maximize overall system efficiency.
Q: Memory prices are skyrocketing, and this situation is unlikely to change in the next two years. From the perspective of CPU or GPU design, does Arm have any relevant technical plans or strategies to cope with the continued pressure of rising memory costs?
Chris Bergey:The mobile industry is indeed currently under cost pressure from multiple aspects, not only memory prices but also wafer costs. Therefore, we are working closely with our partners to promote full-chain optimization, a key aspect of which is driving AI models towards smaller, more efficient directions, achieving better performance and energy efficiency under limited resource conditions. For example, we have collaborated with Alibaba's Tongyi Qianwen to implement a 2-bit quantized model on SME2. Due to the very small model size, it achieves a better balance between performance, cost, and energy efficiency. Furthermore, by enabling graphics capabilities to handle neural network processing, we can alleviate the pressure on other accelerators in the system performing similar tasks, thereby improving efficiency. Thus, we are attempting to address this challenge from multiple different dimensions, but it is indeed a formidable task.
Q: When I attended an Arm event last year, I experienced a phone equipped with the Dimensity 9400 (Cortex-X925 architecture). Later, MediaTek (MTK) released the Dimensity 9500, upgrading the CPU architecture from Cortex-X925 to C1. At the time, I learned that this upgrade would bring significant performance improvements for developers. However, after switching to a Dimensity 9500 phone this year, I did not feel a noticeable improvement in performance or speed in daily use. What factors constrain the difference in actual user experience between the C1 architecture and the previous generation Cortex-X925 architecture? Additionally, after Arm releases the C2 architecture, how will it further accelerate collaboration with partners at the development level to better empower end users?
James McNiven:It is not appropriate for me to comment on the specific product implementations of our partners. Regarding the magnitude of performance improvement, the actual performance depends on the specific workload being run. Upgrading from Cortex-X925 to C1-Ultra has indeed achieved a significant leap in single-thread performance: IPC has seen double-digit growth, and the clock frequency has also increased, resulting in a comprehensive performance improvement of up to approximately 20%, although specific performance will vary depending on the actual workload.
Looking ahead to C2-Ultra and beyond, we will continue to strive for double-digit performance improvements every year. I hope that what you take away from this launch is not just improved benchmark scores, but also continuous improvements in AI performance, app launch speeds, and other user experience metrics.
Regarding Arm's position in the ecosystem: We invest heavily in ecosystem adaptation, promoting support for SME2 across major AI frameworks. Over the past year, we have announced several framework collaborations, implementing SME2 optimizations into products. We also continue to collaborate with the Android ecosystem to improve related efficiency; in the graphics domain, we are working with numerous partners such as Tencent Games and Unity China to implement new features like NSS and Neural Frame Rate Upscaling (NFRU). This is why we are highly focused on improving AI performance on both the CPU and GPU sides. To respond to the question mentioned earlier, such as the neural texture compression mentioned by our partner from Tencent Games today, it is a good example: significantly reducing file size without compromising image quality, saving device memory, which is also a key area we are focusing on.
Q: You mentioned the game "Light and Shadow Reborn" earlier. Could you introduce its current development progress and subsequent release plans?
Chris Bergey:This game is developed by Sumo Digital. We invited them to build this game based on our concepts and neural network technologies, collaborating with them on concept design and other aspects. The game production is now complete. We aim to adapt it to this newly released chip platform, so it will likely require another 30 to 60 days of testing and validation to ensure an excellent experience for players. We will announce the specific release time separately.
Q: Regarding the design of a unified neural network accelerator within the CPU, compared to the traditional independent native AI computing partitioning method of a standalone NPU, what are the differences in computing power utilization? What specific changes have been made at the architectural level?
Chris Bergey:There are considerable differences between the two. The most unique aspect of SME2 is that its programming method is consistent with the CPU: developers do not need to adopt a separate programming model but can continue using familiar CPU programming methods, directly invoking SME2's matrix computation capabilities through instructions. Precisely because of this, SME2 can achieve extremely low latency and is also easier to develop and program.
As you may recall, last year we introduced the Arm Kleidi software library. When developers integrate it into AI frameworks, or use abstracted frameworks like ExecuTorch, the underlying software automatically completes the adaptation. If the hardware is equipped with SME2, it will automatically invoke this capability; if the hardware does not support it, it will not be invoked.
Similarly, the neural network accelerator is programmed via a Vulkan-based GPU workflow. In contrast, NPUs are typically scheduled as independent accelerators, thus adopting a different programming model. In fact, accelerators from different vendors or even different product forms have quite unique programming models. This is precisely what developers need to carefully consider: choosing the appropriate technical path based on the operational requirements of their workloads.
Q: Agents need to continuously call tools and maintain an running state, while smartphones are significantly constrained by battery power consumption and memory resources. How does the Arm platform solve the memory and power consumption challenges brought about by the continuous operation of AI agents?
James McNiven:It depends on how we define the concept of "continuous operation." The phone itself is constantly sending signals to the network and continuously running various tasks in the background. Therefore, there is a distinction between high-intensity occupation of the entire system at all times and periodic scheduling. Currently, many OEMs and agent providers are still exploring implementation schemes for this part. However, achieving efficient continuous operation for agent workloads is not impossible; the key lies in the intensity of the task execution. Some agents can be triggered to run periodically, but they do not require uninterrupted microsecond-level continuous operation.
Chris Bergey:Additionally, I would like to add two points. First, when we play games, the number of simultaneously active threads is typically three to four, and this state may last for an hour or even longer, indicating that intelligent scheduling can enable many possibilities. Second, we can leverage charging scenarios: when you charge your phone at night, the device can perform extensive KV cache optimization and context preprocessing. Therefore, it can be understood as two different operating modes: battery-powered mode and charging-powered mode. Executing high-load tasks such as token generation while charging allows the device to achieve higher operational efficiency during mobile usage phases.
Q: Arm's edge computing layout has currently extended from smartphones to PCs, robotics, and physical AI, among other scenarios, launching corresponding computing subsystems (CSS) for different fields. How does Arm define the boundaries of computing subsystems between different domains? Will this basic CSS subsystem be reused across different domains, or will it continue to evolve with differentiation for specific domains?
Chris Bergey:Simply put, the core judgment criterion is market demand. For example, the Cortex-X925 core used in NVIDIA RTX products is actually a mature technology that was included in CSS for Mobile two years ago. Given the market demand at that time, it was the most suitable and mature solution.
Today, we are seeing stronger momentum in the PC sector, with positive developments not only in the ecosystem for Arm architecture devices but also in other areas. At the same time, we are seeing more and more unique demands, including security capabilities and various other functional features. In such cases, we tailor new CSS and differentiated products accordingly. We invite everyone to continue following our progress. In summary, we reuse mature IPs and components, but the direction of product evolution depends entirely on actual market and customer needs.
Q: In recent years, mobile GPU technology seems to have been following the development path of desktop GPUs. Currently, desktop DLSS 5 has achieved complete frame generation, capable of generating detail elements not rendered by the game itself; whereas mobile GPUs are currently more focused on super-resolution and enhancement. How far is mobile neural graphics technology from reaching the frame generation rendering level of DLSS 5? Does Arm have relevant plans in its technology roadmap?
James McNiven:The underlying technical principles of both are highly similar. We also possess super-resolution capabilities, and we have achieved full frame generation through Neural Frame Rate Upscaling (NFRU). A key difference between our solution and others is that we have placed energy efficiency at the core from the very beginning. This is evident in the model size; constrained by an energy-efficiency-first design philosophy, our NSS model is much more streamlined compared to other solutions with higher complexity. In terms of current capabilities versus DLSS 5—DLSS 5 currently supports multi-frame generation, while NFRU primarily achieves single additional frame generation at this stage, effectively doubling the frame rate. This represents our current technical capability. Of course, we also plan to expand to multi-frame generation in the future.
Furthermore, we will see new application scenarios brought by technologies such as global illumination and radiance caching, which are future technical directions we are laying out. However, at the current stage, we prefer to ensure the efficient implementation of high-quality graphics effects, so our current focus remains on super-resolution and frame rate improvement, specifically achieving a doubling of the frame rate.
One more point worth mentioning is that we have been very actively promoting an open-model ecosystem. Developers can go directly to Hugging Face to download model weights and combine them with the tools we provide for experimentation and development. We hope that this open-source model will stimulate innovation vitality across the industry, including at Arm and other manufacturers. We look forward to more companies fully leveraging this open capability, and we are already seeing increasing industry collaboration. We believe that this model of open collaboration differs from traditional approaches and can bring unique and long-term value to the entire ecosystem.