On July 28, Unity China held the launch event for Harmony Engine 2.0 at the Hyatt Regency Bund Shanghai Hotel, where it also announced a deep strategic partnership with Arm Holdings. The two parties will collaborate to develop super-resolution and frame interpolation features for games based on Arm Neural Technology, bringing mobile gamers PC-like graphical upgrades and frame rate improvements. This collaboration will also effectively lower the barriers for developers in AI graphics development and cross-platform adaptation, accelerating innovation and iteration in mobile gaming.
The biggest highlight of this collaboration is the native integration of Arm Neural Technology—including Neural Super Sampling (NSS), Neural Frame Rate Upscaling (NFRU), and Neural Super Sampling and Denoising (NSSD)—into Unity China's Harmony Engine. As a result, developers do not need to perform separate interface adaptations or learn underlying technical logic from scratch; they can use Arm's AI graphics acceleration capabilities out of the box. This significantly reduces technical learning costs, adaptation and integration costs, and secondary development costs from the source, effectively lowering overall development barriers and accelerating product time-to-market. Meanwhile, leveraging Arm Neural Technology and Harmony Engine's powerful cross-platform native capabilities, this collaboration will also help domestic games achieve higher graphical fidelity and smoother performance across multiple terminals including mobile, desktop, and consoles.
We were also fortunate to interview James McNiven, Vice President of Edge AI Intelligent Terminal Computing at Arm, and Zhang Junbo, CEO of Unity China, along with other domestic media outlets at the event. They shared more details about the collaboration and the technical principles behind Arm Neural Technology. Below is the detailed content of this interview.
Q: If I am a game developer using Unity China's Harmony Engine and want to use Arm Neural Technology features such as Arm Neural Super Sampling (NSS) or Neural Frame Rate Upscaling (NFRU), do I still need to care about the underlying chip situation? In other words, are these features packaged within Harmony Engine so that developers don't need to know or care which GPU supplier the hardware comes from?
Zhang Junbo:Arm Neural Technology is relatively new and will gradually be deployed on flagship devices and more mainstream models in the future. However, the reality for developers is that they usually release only one Android installation package rather than developing a separate package for each specific chip. Therefore, we have designed these features as optional capabilities. Whether they are enabled is automatically determined at runtime based on the device's hardware capabilities, so developers do not need to know in advance whether a specific chip supports related functions. This is a unified solution that performs validation at runtime based on the device's specific configuration.
If other vendors adopt the same API, the handling method remains the same; if APIs are inconsistent, we sometimes choose to support the lowest common denominator or apply different logic depending on the situation. Our goal is always to maintain such a unified approach. In the future, Harmony Engine 2.0 will support more granular profiles, generating differentiated installation packages for different hardware. From a technical perspective, this is a more scientific and refined approach. However, mature supporting solutions have not yet formed in the industry, so we need to further coordinate with channel partners. Technically, we are fully capable of generating differentiated installation packages for different hardware capabilities, but at this stage, we still need to balance distribution efficiency and package management complexity.
Q: Will this fragmented approach have any negative impacts on the overall graphics ecosystem?
James McNiven:I don't think so. Arm Neural Technology will first be deployed on flagship devices and won't immediately cover all devices. In fact, even Arm itself currently offers Arm Accuracy Super Resolution (Arm ASR)—a shader-based super-resolution solution.
For us, what's more important is enabling developers to easily extract optimal performance from available hardware through the engine. Whether the underlying solution is shader-based, neural technology-based, or another hardware acceleration method, the key is helping developers achieve the best results and experience. In the end, the technology that delivers the best results will naturally gain market recognition. The coexistence of different technical routes is normal in industry development and is part of driving continuous improvement in gaming experiences.
Q: Both parties chose to jointly announce this latest collaboration before ChinaJoy. Could you respectively introduce the core positioning of this cooperation, the opportunity for collaboration, and the long-term strategic goals of both sides? We would also like to hear from both of you about the roles Arm and Unity China play in this partnership.
James McNiven:From Arm's perspective, our ultimate goal is for users to have the best gaming experience. To achieve this, developers need to fully leverage hardware performance. Therefore, we hope to engage in productive cooperation with Unity China to make it easier for developers to use neural technology. We will communicate regularly around native integration and technical alignment to ensure smooth implementation. Ultimately, all of this is to enable the technology to be adopted as conveniently and quickly as possible within Unity China's game developer ecosystem.
Zhang Junbo:From the perspective of Harmony Engine, we hope it is not just an alternative option, but a truly leading platform that maintains leadership in graphics rendering capabilities while actively embracing the latest technologies. Arm's main battlefield is mobile, and Harmony Engine is also mobile-centric, so we must support advanced technologies like Arm's in Harmony Engine, which will further enhance our own capabilities.
First, we want to bring these new technologies to developers so they can utilize these new hardware features with the lowest possible cost and speed. Of course, some game companies may choose to call Arm's low-level APIs directly to implement these technologies themselves, but this would be very costly. Rather than having each company repeat the same work, it is better for us to handle this uniformly. This is a very important direction for the entire ecosystem and something we must do.
Q: We see that Arm's neural technology has become dedicated neural accelerators on the GPU. Could you explain specifically where this "dedicated" aspect lies? What are the differences compared to mainstream solutions? For example, what are the core technical route differences compared to NVIDIA's Tensor Core on desktops or other mainstream vendor solutions?
James McNiven:As mentioned earlier, we plan to integrate multiple neural accelerators inside the shader cores of our next-generation GPUs.
Simply put, we have designed a highly efficient neural accelerator. In terms of hardware configuration, it is possible to pair one neural accelerator with each shader core, with the actual deployment quantity adjustable according to hardware configurations. Last year, the Arm Neural Graphics Development Kit also provided early access to ML extensions for Vulkan (VK_ARM_tensors and VK_ARM_data_graph). These extensions allow developers to use tensor resources and build inference pipelines directly into the render graph. This integration helps meet the latency and efficiency targets required by real-time applications on mobile devices. This is not an exclusive solution for Arm; we hope to promote the adoption of these capabilities across the industry, so other GPU vendors can also use these extensions. On this basis, we have also open-sourced the relevant code libraries and algorithms for NSS and NFRU, together forming the complete Arm Neural Technology suite.
In the future, some companies may choose to access these capabilities directly at the API level, while others will invoke them through game engines. We aim to complete the last link in the entire technology stack, enabling developers to more conveniently use the full capabilities from APIs and algorithms to neural accelerators and GPUs, accelerating the popularization and implementation of these innovative technologies.
Q: Can integrating neural technology directly into the GPU architecture be seen as another important technical means and trend beyond improving computing power through traditional process nodes? Also, after applying this neural technology, does it mean developers no longer need to do extensive optimization work for frame rates and power consumption?
James McNiven:We have achieved neural accelerator integration, but this does not mean it replaces traditional technical routes. We still need to continuously improve the performance of traditional rendering; this hasn't changed. Importantly, AI technology can indeed bring higher performance and energy efficiency to more complex content. So I wouldn't call it a "replacement." Developers still need to perform traditional rendering, but on top of that, they can add more processing power to achieve higher resolutions, higher frame rates, and better image quality. Integrating neural accelerators directly into the GPU architecture offers higher energy efficiency compared to deploying them as separate modules, as they can share the memory system with the GPU and work closely with Shader Cores and execution engines.
For developers, they still need to make trade-offs between performance, image quality, and power consumption, which is itself a crucial part of game development. However, neural technology does provide them with more options, allowing for more flexible balancing between different quality and performance targets.
Our goal is to make development simpler. Many games currently in development would struggle to achieve an ideal balance between performance and energy efficiency without the assistance of neural technology. So while Arm Neural Technology is not a "silver bullet" for all problems, it undoubtedly provides significant help to developers.
Q: From Unity China's perspective, could you briefly summarize what kind of trend the integration of game engines and AI represents, and what value it brings? Additionally, what specific practical benefits do Arm's corresponding technologies bring to Harmony Engine, developers, and players?
Zhang Junbo:Overall, AI's revolutionary impact on the gaming industry is mainly concentrated in two core dimensions: content creation and rendering technology, while simultaneously reshaping industry development models and user experiences.
In terms of content creation, developing a game, especially a large-scale one, often doesn't lack creativity or art assets; many times, it's a case of "Everythinggj is ready except for the programmers," because turning creativity into reality requires extensive engineering implementation. In game development, we tend to refer to those responsible for creativity and design as "liberal arts students" and programmers/engineers responsible for implementation as "science students." I personally estimate that the ratio of these two types of talent in the industry is roughly 1:10 or even higher. Turning a good idea into a large amount of usable assets requires strong engineering support.
Game development has always been, in a sense, a tug-of-war between designers, artists, and programmers: artists always want to push for perfection, while programmers say, "We only have this much computing power, so we can only go this far." AI can greatly improve efficiency in this regard—one of the fastest-growing areas for AI in recent years has been programming, and programmers have already significantly boosted their productivity with AI assistance. This has also substantially increased the overall efficiency of game production.
On the other hand, AI progress is also very rapid in areas such as 2D concept art and 3D modeling. It can be said that AI is bringing earth-shaking changes to game productivity. In the future, as long as there is good creativity and sufficient budget to invoke computing power, even small teams may produce high-quality games. Creators who previously could not afford large development teams now have the ability to create well-crafted, creative works through AI.
At the rendering technology level, whether it's ray tracing or traditional rasterization, the basic logic of game engines is to calculate what should be displayed in the next frame based on the current state. This often requires calculating 60 frames per second or more, placing very high demands on computational performance.
This is where super-resolution technology comes into play: for example, we can render images at a lower resolution like 480p and then use super-resolution to upscale them to 1K, 2K, or even 4K. Since computational cost depends on the number of pixels processed, rendering fewer pixels and using neural networks to fill in details allows us to achieve results close to native resolution at a lower computational cost.
Another technology is frame interpolation: due to limited computing power, it's impossible to calculate every frame precisely, but to make the visuals appear smooth, we often need to reach 60 FPS or higher. With AI frame interpolation technology, we can calculate only a portion of the frames, such as 30, and let AI generate the remaining frames, thereby achieving a smoother visual experience under limited computing power.
In the future, there may be "world models" trained specifically for games, turning game logic into a foundational model that ultimately reconstructs and generates the visuals and content players see in-game through this world model. I believe this could be a trend. Such "world models" will bring more possibilities to games: for instance, NPC behavior, dialogue, and even animation performance could be dynamically generated on-device based on players' real-time actions, offering greater diversity. These are the changes we believe AI will bring to gaming.
Looking further ahead, will AI play games for players? Honestly, for players like me who aren't very skilled, I can't beat others in many action games. But if in the future I could "hire" an AI agent to battle for me, perhaps the determinant of a "world champion" would no longer be the player's own skill level, but rather the strength of the AI computing power they invoke. For average players like us, this might be a promising future. I look forward to sending my own AI to fight for me.
Overall, the gaming industry has always been one of the fastest drivers of computer technology and computer graphics development, with many technological innovations pioneered by the gaming sector.