I played Android’s first AI frame-gen game: Mobile just had its DLSS moment

Talk of the mobile silicon town this year is, once again, AI, but this time with a twist for gamers. The latest GPU developments from Arm and Qualcomm center on similar neural upscaling techniques we’ve long seen in the PC space, using “AI” to boost frame rates and resolutions, making next-gen games look and run better without the traditional rendering costs.

In a bid to showcase that this is more than just tech demo talk, Arm worked closely with developer Sumo Digital to create Neural Dawn — a light platforming title built from the ground up to use Unreal Engine’s impressive MegaLights ray-traced lighting system. The game leverages Arm’s Neural Super Sampling and Denoising (NSSD) and Neural Frame Rate Upscaling (NFRU) technologies to make this actually run on a smartphone. Version 1.0 just dropped last week, meaning we can actually go and play it.

As you might expect from a bleeding-edge title, Neural Dawn has some rather strict system requirements. You need an Arm Mali G2 Ultra NX GPU to run it, making the MediaTek Dimensity 9600 Pro the only chipset on the market that currently supports the necessary hardware. I grabbed my OPPO Find X10 Pro Max, downloaded the APK, and booted up the game.

I can tell you off the bat that the lighting effects are undeniably impressive. The opening scene features tens, if not hundreds, of real-time lights casting dynamic shadows across the environment without a hiccup. This is the closest I’ve seen to genuine AAA-quality graphics on a smartphone, but just how did they do it?

How can you ray-trace lights on mobile?

I won’t delve too deep, but let’s quickly cover what MegaLights, NSSD, and NFRU actually do.

Let’s start with NFRU, which boosts game smoothness. Essentially, Arm uses an AI neural network running on the GPU to generate brand-new “in-between” frames. For every frame the game engine natively renders, NFRU creates an interpolated twin. This means that even if a heavy game is only running at 30 frames per second (FPS), you get a beautifully smooth 60 FPS output on your screen.

NSSD is the partner technology, but it focuses on image quality instead of smoothness. It takes a game running at a low resolution and upscales it to look sharp and pixel-dense. This isn’t just stretching the image; NSSD studies motion, depth, and data from previous frames to intelligently fill in the blanks. At the same time, NSSD cleans up the jagged edges of objects (anti-aliasing) and wipes away visual noise from advanced ray-traced lighting.

Doing all of this at once is incredibly important for features like Unreal’s MegaLights. To calculate hundreds of lights simultaneously, MegaLights uses a mathematical shortcut that fires random, scattered rays at the highest priority light. Without a cleanup process, this raw lighting looks like a boiling, flickering mess of static across your screen. NSSD’s built-in “denoiser” acts as a filter, instantly smoothing out that static into clean, realistic lighting in a single step.

In previous technical discussions for Neural Dawn, Arm highlighted taking a low 540p baseline image and scaling it up to 1080p. This low starting resolution is the secret to how the phone can handle so many complex light sources at once. While the final result looks a little softer than a native 1080p image, on a small, sub-7-inch smartphone display, I found it very hard to tell the difference. Aesthetically, the results are undeniably impressive; check out my screenshots below.

Neural Dawn performance results

In terms of performance, Neural Dawn runs pretty flawlessly on the OPPO Find X10 Pro Max. I recorded a healthy mean frames per second (FPS) of 59.8, with 5% lows at 58.6 and 1% lows at 57.2 FPS. In other words, it is tightly locked to the maximum 60 FPS target. However, given that the title uses NFRU to insert a “fake” frame every other frame, the underlying engine is presumably running closer to 30 FPS before upscaling the result. From my playthrough, the experience did indeed feel incredibly smooth, and I couldn’t easily tell that the game wasn’t running natively at this speed.

I played the game for just over 30 minutes straight. The detailed performance graph above shows a slight dip after the first five minutes of play. Neural Dawn was initially locked to a very tight 60 FPS, but once the phone warmed up slightly and hit some more open-ended areas, the frame rate began to fluctuate due to the occasional missed frame. However, it should be noted that the graph plots close to 20 minutes of data, and those occasional missed frames land at a pretty steady 33ms — meaning it is far from a janky experience. The only noticeable frame drops occurred during early menu transitions and loading screens.

Fortunately, there is no sign of severe performance throttling or a sudden spike in heat. Temperatures gradually peaked at a just-noticeably-warm 37.7°C after around 18 minutes of play and held steady alongside performance until the end of the session. That’s a modest ramp-up and safe peak temperature, making Neural Dawn run far cooler than I’ve seen other phones handle benchmarking stress tests. For a technological showcase, Arm and Sumo Digital successfully pulled off a great-looking and well-performing piece of software.

Is this the future of mobile gaming?

What impressed me most about Neural Dawn isn’t actually the graphics, even though they look great for a mobile title. Instead, the rock-solid frame rates, the lack of major hitches, and the fact that it doesn’t obliterate the phone’s battery or bake your hands should be the big talking points. Neural Dawn doesn’t push graphical fidelity by putting extreme strain on the hardware; it achieves this within the exact same performance and power budgets that popular mobile games already live with.

That being said, I did experience a few issues that gave me pause for thought about whether this is a technology I want in every mobile title.

The most common and noticeable issue I encountered was motion ghosting around the player character and some world objects while turning the camera. The result is a shimmering, almost blocky halo during fast camera panning, presumably caused by a combination of NSSD and NFRU lacking enough data to accurately fill in the edges around rapid motion. The effect isn’t terrible enough to ruin the experience, but once you notice it, it is hard to unsee. Thankfully, it was only jarring while spinning the camera around; I didn’t notice the same artifacts while moving the character forward through the world or observing more distant objects in motion. The images below capture the issue, though it is admittedly less noticeable in motion.

Despite the frame-rate counter locking tightly to 60 FPS, the game also didn’t always feel quite as buttery smooth as the numbers suggest. For instance, some particle animations appeared notably slower and jankier than the surrounding gameplay motion, which was a little jarring. Similarly, I felt a handful of instances where control latency became disconnected from the action on screen. I could sense a slightly sluggish responsiveness at rare points, despite the visuals remaining silky smooth.

These are typical complaints with frame upscaling technology, especially when taking a relatively low baseline frame rate or resolution and pushing it to a more attractive output. Going from 30 FPS to 60 FPS can be cutting it fine; a missed frame or two above 33ms can feel sluggish, whereas taking 60 FPS to 120 FPS offers far more headroom before a game feels janky to play. Likewise, using upscaling and denoising to quadruple a game’s rendering resolution from 540p to 1080p leaves far more room for smudging, blurring, and other artifacts compared to taking an already healthy 1080p up to 4K.

This isn’t to disparage Arm’s upscaling technologies. Arm has clearly succeeded in taking graphical effects that would otherwise struggle to run on a mobile GPU and making them not only playable, but smooth — all within a power budget that won’t cook your handset. But there is no such thing as a free lunch. What Neural Dawn gains in graphical fidelity, frame rate, and power efficiency, it trades off with small, occasionally noticeable AI upscaling artifacts. The good news is that these technologies can and will continue to improve in subsequent generations.

So, is this the future of mobile gaming? Well, I think so — at least in part.

As ever, mobile’s biggest problem is the classic chicken-and-egg conundrum. Ray-tracing and AI upscaling remain a technological niche reserved for the cutting-edge and most expensive smartphones on the market. The OPPO Find X10 Pro Max I played Neural Dawn on costs a whopping £1,799; you could buy three PlayStation 5s for that!

At the same time, developers naturally want their titles to run on every “potato” phone out there to maximize their player base and potential revenue. Adopting these high-end technologies is likely lower down their priority list than optimizing games to run on mid-range hardware.

That doesn’t mean AI upscaling won’t carve out a permanent place for itself in mobile gaming. In fact, my playthrough has convinced me that this is almost certainly mobile’s best shot at achieving console- and PC-quality graphical fidelity in the palm of our hands. If we want to see AAA games ported like-for-like onto phones, upscaling is the key. However, whether or not ideas like Arm’s NFRU or Qualcomm’s equivalents end up in the majority of next-gen mobile games is something we will just have to wait and see.