Arm Details Its Mobile‑First AI Reconstruction Approach, Diverging 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 mobile graphics. Central to the announcement was a new AI‑centric compute platform, CSS for Mobile 2, and the Arm Mali G2‑Ultra NX graphics processor. The G2‑Ultra NX is built around neural‑computing primitives, promising higher visual fidelity while dramatically cutting heat output and power draw compared with conventional GPU pipelines.

Arm positions its on‑device AI solution as a more transparent alternative to Nvidia's controversial DLSS 5. Rather than relying on a closed, proprietary model, Arm’s approach is open‑source, allowing developers to inspect, modify, and even retrain the underlying neural networks. This openness aligns with the original intent of many game creators, who want control over how their artistic style is reproduced during up‑scaling and frame‑generation.

While Arm did not disclose specific product partners for the Mali G2‑Ultra NX, its predecessor, the Mali G1‑Ultra, already powers smartphones from Xiaomi, Vivo and OnePlus. Deyan Lazarov, a member of Arm’s GPU team within the Edge AI division, explained that the new chip has been in development for roughly two and a half years. Its primary goal is to give mobile developers the ability to employ compute‑heavy techniques—most notably ray tracing—without overwhelming the limited thermal and battery budgets of handheld devices.

"We’re seeing a steady rise in ray‑tracing adoption on mobile titles," Lazarov noted. "It used to be a niche, but more games are now looking to add realistic lighting and reflections, and the G2‑Ultra NX is designed to make that feasible." The GPU’s capabilities were demonstrated in a collaboration with Sumo Digital on the demo title *Neural Dawn*, which showcases how the chip can handle Unreal Engine features such as MegaLights. Other upcoming games slated to leverage the hardware include Tencent’s *Arena Breakout Infinite*, NetEase’s *Where Winds Meet*, and Infold Games’ *Infinity Nikki*. Lazarov emphasized the unique constraints of smartphones: unlike consoles or PCs, phones cannot dissipate large amounts of heat and must run continuously on battery power.

To address this, Arm integrates neural processing directly into the rendering pipeline. By off‑loading portions of the image synthesis to dedicated neural accelerators, the GPU workload shrinks, leading to lower power consumption and, consequently, a more efficient overall system. "In practice, this means developers face fewer trade‑offs," Lazarov said.

"They can allocate the saved budget to higher visual fidelity, smoother frame rates, longer battery life, or any combination that fits their vision." Arm backs this promise with an open‑software strategy and a toolbox of developer resources. Over the past two years, the company has worked closely with the ecosystem to produce AI‑native graphics that are both powerful and developer‑friendly. The suite includes Neural Super Sampling (NSS), Neural Super Sampling and Denoising (NSSD), and Neural Frame‑Rate Upscaling (NFRU). These technologies enable the system to upscale lower‑resolution frames, add extra rays for richer lighting, and generate intermediate frames using AI reconstruction.

Lazarov illustrated the workflow with a simple example: in one scenario, the GPU renders only a quarter of the final image, while the remaining three‑quarters are reconstructed by the neural accelerators via NSS. In a more extreme case, the entire frame is produced by the AI engine, with the GPU performing virtually no traditional rendering work. This flexibility allows developers to tailor the balance between native GPU rendering and AI‑assisted reconstruction based on performance targets. Comparisons with Nvidia’s DLSS 5 are inevitable.

While DLSS adds detail through generative models, Arm’s networks are strictly convolutional, meaning they do not hallucinate new content but rather learn to recognize and reproduce the patterns present in real training data. All of Arm’s networks—resolution up‑scalers, frame‑rate generators, and others—are publicly available on GitHub and Hugging Face, enabling developers to fine‑tune them with bespoke data sets that reflect a game’s unique art style.

"If you have a quirky visual direction, you can capture data from your own game, retrain our network, and ensure the AI up‑scales exactly as you intend," Lazarov explained. "There’s no risk of losing IP control because you own the retraining process." By contrast, DLSS is a closed‑source black box: developers cannot see the training data, nor can they retrain the model to suit their specific aesthetic.

It operates on a take‑it‑or‑leave‑it basis, which many creators find limiting. Another key distinction is mobile‑first design. Arm’s neural networks are built with strict power, bandwidth, and memory constraints in mind.

Attempting to run a desktop‑oriented solution like DLSS on a phone would instantly exceed those limits, whereas Arm’s solution stays within the tight budgets of handheld hardware. Lazarov highlighted impressive efficiency gains: the G2‑Ultra NX delivers up to four times the performance and four times the efficiency of its predecessor, while slashing DRAM traffic by roughly 70 %. Reducing memory accesses is crucial because DRAM reads are a major source of power draw on mobile chips.

In a head‑to‑head test, the older GPU could only sustain *Neural Dawn* at about 40 fps when pushed to its limits, dropping to 15 fps under normal conditions. The G2‑Ultra NX, however, comfortably maintains 60 fps, demonstrating the practical impact of the AI‑enhanced pipeline.

On the tooling side, Arm currently offers plug‑ins for Unreal Engine and plans to expand support to Unity and other engines. A recent partnership with Unity China signals that broader engine integration is on the roadmap. For studios using custom engines, Arm provides an SDK that streamlines integration without requiring engine‑specific plug‑ins. Lazarov stressed that developers can retrofit existing titles with Arm’s AI graphics in as little as three months, delivering immediate visual and performance upgrades.

The *Neural Dawn* demo itself was built by a team of fewer than 20 people in under 18 months, showcasing how Arm’s technology can accelerate development cycles while maintaining high quality. In summary, Arm’s mobile‑first AI reconstruction strategy offers an open, efficient, and adaptable alternative to the proprietary DLSS 5 model.

By embedding neural accelerators into the graphics pipeline, providing publicly accessible networks, and delivering robust developer tools, Arm aims to empower creators to push mobile gaming visuals toward console‑level fidelity without sacrificing battery life or thermal headroom.