Arm Details Its Mobile‑First AI Reconstruction Approach, Offering an Alternative to the "Black Box" DLSS 5 Model

Arm, the British chip and software design powerhouse, staged its Arm Everywhere China showcase in Shanghai this week, unveiling a suite of fresh AI‑centric technologies aimed at reshaping mobile graphics. The centerpiece of the announcement was a new AI‑native compute platform called CSS for Mobile 2 and the Mali G2‑Ultra NX graphics processor, a GPU built around neural‑accelerated rendering.

Arm says the combination is designed to lift visual fidelity on phones, tablets, laptops and even handheld consoles while keeping power draw and heat generation to a minimum. While Arm did not name specific products that will ship with the Mali G2‑Ultra NX, the earlier Mali G1‑Ultra has already found its way into smartphones from brands such as Xiaomi, Vivo and OnePlus, suggesting that the next‑generation chip could appear in a similar range of devices later this year. Deyan Lazarov, a senior member of Arm’s GPU product team within the Edge AI division, explained that development on the G2‑Ultra NX has spanned roughly two and a half years.

The goal, he said, was to give mobile developers a practical path to incorporate compute‑heavy techniques—most notably ray tracing—without sacrificing the battery life or thermal envelope that mobile platforms demand. "There’s been a growing appetite for ray tracing on mobile titles," Lazarov noted. "It started out slow, but now more games are actively looking to enable it, and our hardware is built to support that trend." The new GPU, according to a GamesIndustry.biz report earlier in the year, can tap into Unreal Engine features such as MegaLights, and Arm has already partnered with Sumo Digital to produce a demo title called *Neural Dawn* that showcases the chip’s capabilities.

Lazarov described the target audience as "mobile‑first games that still want the visual polish of desktop‑class titles". He added that many modern games are designed to be cross‑platform, meaning players expect a seamless experience whether they are on a home console, a PC, or a phone on a train. "Developers want a unified experience that doesn’t degrade dramatically when you move from a powerful console to a battery‑powered device," he said. A handful of upcoming games have already signed up to use the Mali G2‑Ultra NX, including Tencent’s *Arena Breakout Infinite*, NetEase’s *Where Winds Meet* and Infold Games’ *Infinity Nikki*.

Lazarov emphasized that mobile hardware faces unique constraints: unlike consoles, phones cannot dissipate large amounts of heat and must run continuously off a battery. To address these limits, Arm is blending neural processing with the traditional graphics pipeline.

By offloading portions of the rendering workload to dedicated neural accelerators, the GPU can render fewer native pixels while the AI system reconstructs the missing data, dramatically lowering power consumption and thermal output. "In practice this means you get a new level of efficiency that eases many of the trade‑offs developers normally have to make," Lazarov explained.

"They can choose to reinvest the saved budget into higher resolution, smoother frame rates, longer battery life, or a combination of all three." Arm is backing this vision with an open‑source software stack. Over the past two years the company has been collaborating with the broader ecosystem to make its AI tools developer‑friendly. The core technologies include Neural Super Sampling (NSS), Neural Super Sampling and Denoising (NSSD) and Neural Frame Rate Upscaling (NFRU). Together they allow the system to upscale images, add extra ray samples and even generate entirely new frames via AI‑driven reconstruction.

Lazarov illustrated the process with a two‑frame example. In the first frame, the GPU renders only 25 % of the image; the remaining three‑quarters are filled in by the neural accelerator using NSS.

In the second frame, the GPU does no traditional rendering at all—the neural engine creates every pixel, delivering a fully AI‑reconstructed frame. This approach contrasts sharply with Nvidia’s DLSS 5, which has sparked debate because its generative networks can sometimes diverge from a developer’s artistic intent.

"Our networks are not generative; they’re convolutional, so they don’t invent new content—they simply know how the final image should look based on real‑world training data," Lazarov said. "We trained them on authentic content and published the models on GitHub and Hugging Face, making them openly accessible." Because the models are open, developers can fine‑tune them with their own art assets. If a game has a distinctive visual style, a studio can capture data from that game, retrain the network, and ensure that upscaling and frame generation respect the original aesthetic. This level of control eliminates concerns about losing intellectual‑property rights or having a black‑box system impose unwanted changes.

By comparison, DLSS is a closed, proprietary solution. "You can’t see what it’s been trained on, you can’t retrain it, and you’re stuck with a take‑it‑or‑leave‑it proposition," Lazarov argued.

"Our approach lets you take the model, adapt it, and make it work exactly the way you need it to." Another key distinction is the mobile‑first design philosophy. Arm’s networks are built with strict power, bandwidth and memory budgets in mind.

DLSS, even if ported to a phone, would instantly exceed those limits, according to Lazarov. The Mali G2‑Ultra NX therefore achieves impressive efficiency gains: four‑times the performance, four‑times the efficiency, and a 70 % reduction in DRAM traffic—a major source of power draw on handheld devices. To illustrate the performance uplift, Lazarov cited a benchmark with *Neural Dawn*. On the previous generation GPU, the game capped at around 40 fps under ideal conditions and dropped to 15 fps in typical play.

On the new Mali G2‑Ultra NX, the same title runs smoothly at 60 fps, demonstrating how AI‑assisted rendering can unlock higher frame rates without sacrificing visual quality. Arm is also investing in developer tooling. Plug‑ins for Unreal Engine are already available, and a Unity integration is on the roadmap, following a recent partnership with Unity China.

For studios that rely on custom engines, Arm offers an SDK that provides a streamlined path to integrate its AI‑native graphics pipeline. Lazarov highlighted that retrofitting an existing shipped game with Arm’s solution can be done in as little as three months, allowing developers to quickly deliver visual upgrades or efficiency improvements to their user base. He pointed out that *Neural Dawn* itself was built by a team of fewer than 20 people in under 18 months, underscoring how Arm’s technology can accelerate development cycles and reduce the resources required to produce cutting‑edge mobile experiences.

In summary, Arm’s mobile‑first AI reconstruction strategy presents a transparent, open, and power‑aware alternative to Nvidia’s DLSS 5. By publishing its convolutional networks, offering flexible retraining, and tightly integrating neural accelerators with the graphics pipeline, Arm aims to give developers the tools they need to deliver console‑level visuals on battery‑powered devices without compromising on performance, battery life or artistic intent.