Arm Details Its Mobile-First AI Reconstruction Approach, Distinct from the "Black Box" Model of DLSS 5

Arm, the British semiconductor and software design powerhouse, recently staged its Arm Everywhere China showcase in Shanghai, unveiling a suite of fresh AI‑computing innovations aimed at reshaping the mobile graphics landscape. Central to the announcement was the introduction of an AI‑native compute platform dubbed CSS for Mobile 2, alongside the brand‑new Arm Mali G2‑Ultra NX GPU. This graphics processor is built around neural‑accelerated rendering, promising a notable lift in visual fidelity while dramatically curbing heat output and power draw compared with conventional mobile GPUs.

Arm positions its on‑device AI solution as a more transparent alternative to Nvidia’s controversial DLSS 5, arguing that the latter’s closed‑source, generative models stray from developers’ artistic intent. Instead, Arm’s approach relies on convolution‑type networks that simply learn how existing content should appear, enabling upscaling and frame generation without inventing new visual elements.

All of these networks—resolution upscalers, frame‑rate generators, denoising modules—are publicly available on GitHub and Hugging Face, allowing developers to fine‑tune them to the unique aesthetic of their games. While Arm did not disclose specific product partners, the predecessor Mali G1‑Ultra chip already powers devices from Xiaomi, Vivo and OnePlus, suggesting that the G2‑Ultra NX could soon appear in a similar range of flagship smartphones, tablets, laptops and even handheld consoles.

Deyan Lazarov, a senior member of Arm’s GPU product team within the Edge AI division, explained that the G2‑Ultra NX has been in development for roughly two and a half years. Its primary goal is to give mobile creators the ability to harness compute‑heavy techniques—most notably ray tracing—without sacrificing battery life or thermal headroom.

"We’re seeing a growing appetite for ray tracing on phones," Lazarov noted. "It took a while to gain traction, but more titles are now looking to incorporate realistic lighting and reflections." The new GPU integrates tightly with Unreal Engine, supporting advanced features such as MegaLights. Arm collaborated with Sumo Digital to produce a demo title, *Neural Dawn*, which showcases the G2‑Ultra NX’s capability to render high‑quality scenes at smooth frame rates. Other upcoming titles slated to benefit from the chip include Tencent’s *Arena Breakout Infinite*, NetEase’s *Where Winds Meet*, and Infold Games’ *Infinity Nikki*.

Lazarov emphasized the distinct constraints of mobile hardware: unlike consoles or PCs, phones cannot afford large heat sinks or continuous power from the grid. "A mobile device runs on a battery and has a much tighter thermal envelope," he explained. "Our solution blends neural processing with the traditional rendering pipeline, offloading a substantial portion of the GPU workload to dedicated neuro‑accelerators. This reduces both GPU power consumption and overall system energy use." The practical upshot, according to Arm, is that developers no longer need to make as many compromises between visual quality, performance and battery longevity.

By freeing up compute budget through AI‑driven efficiency gains, studios can choose to reinvest those savings into sharper textures, higher frame rates, longer play sessions, or a mix of all three. Arm’s AI‑native graphics stack comprises three core technologies: 1.

**Neural Super Sampling (NSS)** – an upscaling algorithm that reconstructs missing pixels using learned patterns. 2. **Neural Super Sampling and Denoising (NSSD)** – combines upscaling with noise reduction for cleaner images.

3. **Neural Frame Rate Upscaling (NFRU)** – generates intermediate frames to boost perceived frame rates.

In practice, NSS can render only a quarter of a frame on the GPU, with the remaining three‑quarters synthesized by the neuro‑accelerators. In an even more extreme scenario, the entire frame may be produced solely by AI, eliminating traditional rasterization for that moment.

This contrasts sharply with Nvidia’s DLSS 5, which relies on generative networks that can introduce details not present in the source content, sometimes diverging from the creator’s vision. Lazarov clarified the philosophical difference: "Our networks are not generative; they don’t imagine new content. They have been trained on real‑world data and are open‑source, so developers can retrain them on their own art style. This guarantees that the upscaled or interpolated frames stay true to the game’s original look, preserving intellectual property and artistic intent." By contrast, DLSS is described as a proprietary black box: developers receive a pre‑trained model with no insight into its training data or internal mechanics, and they cannot modify it to suit a unique visual style.

Arm argues that this closed approach limits creative control and can lead to unintended visual artifacts. Another key distinction is mobile suitability. DLSS, even if ported, would quickly exceed the power, memory bandwidth, and thermal budgets of a smartphone. Arm’s networks, however, have been engineered from the ground up to operate within the tight constraints of mobile silicon, balancing performance with power efficiency.

Quantitative claims from Arm include up to a four‑fold increase in both performance and efficiency, alongside a 70 % reduction in DRAM traffic—a major source of power consumption on handheld devices. Lazarov illustrated the impact with a head‑to‑head: the previous‑generation Mali GPU could only sustain *Neural Dawn* at roughly 40 fps under ideal conditions and dropped to 15 fps in typical use, whereas the G2‑Ultra NX comfortably hits 60 fps.

To support developers, Arm has released Unreal Engine plug‑ins and is working on expanding to Unity and other engines. An SDK provides a streamlined integration path for custom engines, ensuring that even studios without dedicated plug‑ins can adopt the technology.

Arm also highlighted a partnership with Unity China, signaling future Unity support. Lazarov stressed that integrating Arm’s AI graphics into an existing title can be rapid. "A shipped game can be updated with our solution in as little as three months," he said, allowing studios to refresh legacy titles with modern visual enhancements without a full remake.

The *Neural Dawn* demo itself was built by a small team of fewer than twenty developers in under 18 months, underscoring how Arm’s tools can accelerate development cycles. By offloading heavy rendering tasks to AI, developers can focus on gameplay and design rather than spending months optimizing low‑level graphics code. In summary, Arm’s mobile‑first AI reconstruction strategy offers an open, developer‑friendly alternative to Nvidia’s DLSS 5.

By leveraging convolutional networks that are publicly available and retrainable, the company aims to give creators full control over how their games look on mobile devices, while delivering substantial gains in performance, power efficiency, and memory usage. This approach could reshape how high‑fidelity graphics are delivered on smartphones, making console‑level visuals more attainable on the go.