Qualcomm unveils Adreno Neural Fusion – an analog of DLSS for smartphones
Qualcomm has unveiled Adreno Neural Fusion, a new neural rendering system for mobile games. It combines AI-based image scaling and intermediate frame generation, and Qualcomm claims it reduces rendering power consumption by up to 40% compared to the company’s previous solution.
How Adreno Neural Fusion Works
Adreno Neural Fusion consists of two core technologies: Super Resolution and Frame Generation. The first allows the game to render an image at a lower resolution, after which the neural network restores it to the target resolution. The second creates an additional frame between two frames that were actually rendered by the game engine.
When both features are used simultaneously, the game may render images at half resolution and calculate half the actual frame count. The system then outputs a full-resolution image with double the frame rate relative to what the engine sends.
Super Resolution operates on a temporal principle and uses data from multiple consecutive frames. The algorithm receives color images, scene depth, motion vectors, and slight camera shifts between frames. This enables the recovery of fine details and reduces flickering in thin objects.
Frame Generation uses two consecutive full-resolution frames, depth data, and motion vectors to compute an intermediate scene state. In the first version of SDK, one additional frame is generated for every rendered frame. Qualcomm explicitly notes this results in increased latency by approximately the duration of one real frame.
AI now works directly inside GPU
The main hardware change appears in the Snapdragon 8 Elite Extreme Gen 6. GPU now includes three Adreno Matrix Cores—one dedicated AI block in each of the three graphics slices. They perform matrix operations directly within the graphics pipeline, so data no longer needs to be constantly transferred between GPU and a separate neural processor.
Matrix Core is backed by 18 MB of Adreno High Performance Memory. This memory, located adjacent to GPU, stores intermediate neural network results, frame buffers, and other data that would otherwise need to be transferred to system LPDDR. Qualcomm expects this to reduce latency and power consumption.
In its own test, Dragon Alley reported a reduction in rendering power requirements of up to 40% compared to the previous solution. This is Qualcomm’s result; independent tests of Adreno Neural Fusion on commercial smartphones are not yet available.
The Matrix Core hardware is a feature of the Snapdragon 8 Elite Extreme Gen 6, but the documentation for the new SDK describes broader compatibility. Qualcomm specifies support for Adreno Neural Fusion starting with Snapdragon 8 Elite Gen 6. On devices without Matrix Core, computations can be redirected to NPU via the GPU-NPU pair, and the optimal execution path is selected automatically.
Support for Unreal Engine and Unity
Qualcomm released Adreno Neural Fusion SDK for developers. It includes tools for Vulkan, plugins for Unreal Engine 5, support for Unity 6.6 and later, as well as integration with Snapdragon Profiler.
SDK also includes debugging tools that allow separate viewing of motion vectors, scene depth, reprojection, and data accumulation between frames. This should simplify finding common neural scaling issues such as trails, incorrect reconstruction of moving objects, and image flickering.
Support for the technology is implemented at the level of a specific game, so it is not possible to force-enable Adreno Neural Fusion in any Android project.
Qualcomm states it is working with more than 20 games. The named projects include Honkai: Star Rail, War Thunder Mobile, Diablo Immortal, Arena Breakout, Farming Simulator 26, Monster Hunter Outlanders, Neverness to Everness, and Where Winds Meet.
The main benefit of the technology for smartphones lies not only in increasing the visible frame rate. Reducing internal resolution lowers the load on GPU and memory bandwidth, which should slow device heating and allow the graphics processor to maintain high frequencies for longer. How noticeable this effect will be in real games will become clear after the first mass-produced smartphones with Snapdragon 8 Elite Extreme Gen 6 appear.
