At GDC 2026 on March 10, 2026, Qualcomm disclosed a development codename—"Adreno AI Frame Fusion (AIFF)"—combining neural super-resolution with frame generation. When Notebookcheck reported on July 26 based on internal documents, a concrete hardware picture emerged: a dedicated matrix compute unit on an unannounced SoC called "SM8975," differentiating features from a standard variant.
However, what Qualcomm officially explained extends only to the technology under development and its integration into game engines. The product name "Snapdragon 8 Elite Gen 6 Pro," the Adreno 850, two Matrix ALUs, and the Pro-exclusive segmentation have all not been confirmed by the company. What's confirmed is the technical overview of AIFF; the product configuration lacks official announcement.
AI Frame Fusion Already Appeared in March
In its GDC talk, Qualcomm described AIFF as a combination of "2x upscaling" and "2x frame generation." The examples given were rendering a game at 960×540 pixels and upscaling to 1920×1080 pixels, and generating one frame between frames rendered at 30 frames per second to display at 60 frames per second. The former reduces the rendering load per frame, while the latter smooths perceived motion on screen.
This represents a major structural shift from Snapdragon Game Super Resolution (GSR), which Qualcomm released in 2023. The original GSR was a spatial-type technique that referenced only a single low-resolution frame, processing upscaling and edge enhancement in one pass. AIFF's super-resolution is temporal, passing low-resolution color, depth, motion vectors, and jitter amount, along with the previous high-resolution output and feature data, to the next frame. A four-layer convolutional network uses both past and current information, aiming to reproduce fine shapes and moving particles accurately.
Frame generation draws on the same category of information to create one frame between already-rendered frames. Qualcomm kept the algorithm's details undisclosed, but did indicate a design where frame upscaling and intermediate frame generation can be called separately from the SDK as distinct processes. Games can choose whichever they need.
The official explanation contains no specification for "upscaling from 1080p to 1440p." The concrete example Qualcomm presented was 2x upscaling from 540p to 1080p, while the July report that started this story states the leaked document mentions output modes of either 1080p or 1440p. Confusing input resolution with output resolution changes both the required computational load and image quality assessments.
18MB HPM and DRAM Bandwidth
AIFF's convolutional processing places intermediate results in Adreno High Performance Memory (HPM). According to Qualcomm, the goal of using HPM is to reduce the bandwidth and power consumed by shuttling data back and forth to DRAM outside the SoC. The current Snapdragon 8 Elite Gen 5 already includes 18MB of HPM, and the company claims up to 10% power efficiency improvement and up to 38% performance gains in games.
In neural rendering, where data is stored matters as much as how fast the compute units run. Reading 960×540 color, depth, and motion buffers while updating high-resolution history and feature data every frame increases transfers with external memory. For smartphones, the harder challenge isn't peak performance but avoiding clock throttling from heat during extended play sessions. Qualcomm itself cited power consumption and thermal throttling as the starting point for AIFF development at GDC.
On the software side, Vulkan serves as the interface between game engines and the AIFF SDK. Khronos's "VK_QCOM_data_graph_model" is an extension specification for running QNN-format models on Adreno GPUs or Hexagon NPUs while sharing memory with GPU image processing. Qualcomm's talk also described a design where buffers aren't copied between the GPU and machine learning compute units, instead using Vulkan's synchronization mechanisms for processing. This proposal document was last updated on June 24, 2025, and remains unratified as of now.
The Pro-Exclusive Claim Surrounding Two Matrix ALUs
Notebookcheck, which reviewed the leaked document, reported that the SM8975 features an Adreno 850 with two Matrix ALUs built into its shader processors. These units, handling matrix multiplication and convolution, are placed near the GPU and combined with 18MB of GMEM to run AIFF. Meanwhile, the standard variant, reportedly the SM8950, uses an Adreno 845 with 12MB of GMEM and lacks Matrix ALUs, meaning it wouldn't support AIFF either.
If this differentiation proves accurate, Qualcomm would be making neural rendering itself—alongside differences in CPU clock speed and GPU scale—a reason to buy the Pro variant. Since the current Snapdragon 8 Elite Gen 5 already includes 18MB of HPM, the novelty of the next-generation Pro lies not in memory capacity but in the addition of Matrix ALUs and SDK utilization. The report that the standard variant would shrink to 12MB could also affect what resolution and image quality the same games can maintain.
That said, Qualcomm's current product page lists the Snapdragon 8 Elite Gen 5 as its latest flagship. Model numbers like SM8975, SM8950, Adreno 850, and Adreno 845, as well as the Pro/standard split, remain unannounced. Notebookcheck presents this as independent reporting based on internal documents, but it doesn't substitute for an official spec sheet.
Not a One-Tap Feature—Game-Side Implementation Is Required
The AIFF that Qualcomm presented at GDC isn't a feature that the OS or drivers can retroactively apply to any video content. Games must pass low-resolution color, depth, motion vectors, and jitter amount to the SDK to execute super-resolution processing. For frame generation, a high-resolution image after post-processing but before UI rendering is also required. Both the generated frame and normally rendered frames need UI overlaid separately, and display timing must be adjusted accordingly.
For this reason, Qualcomm plans to provide a complete package including plugins for Unreal Engine and Unity, Snapdragon Profiler support, samples, and documentation. Unreal Engine 5 can use the existing interface for temporal upscalers, though versions before 5.6 require small engine modifications for jitter handling. Bundling the SDK with each game and integrating it title by title is the currently disclosed path.
Adding a universal switch for emulators or existing games would require a separate layer to capture depth and motion information that games normally output, plus a mechanism to isolate the UI. Qualcomm has not announced providing this. Even if a smartphone is chosen solely for its SoC, the feature won't work unless the game itself implements AIFF.
The example of displaying 60 frames per second from 30 frames per second also has limitations. Since the added frame doesn't contain new game logic or input results, while the visuals appear smoother, the update frequency of controls doesn't actually increase to 60 times per second. Qualcomm has not yet disclosed AIFF's latency, processing time, or power consumption, and even acknowledged that the 1080p demo at GDC contained artifacts visible upon extended viewing.
A 2.3ms Prior Study and Numbers Still Missing as of September
"Mob-FGSR," presented at SIGGRAPH 2024, reported generating frames and high-resolution images from color, depth, and motion vectors, processing 720p frame generation in 2.2ms and 1080p frame generation plus super-resolution in 2.3ms on a Snapdragon 8 Gen 3. This was a research implementation that didn't use a neural network at runtime, differing from AIFF's approach, but it demonstrated two years ago that both processes could be combined on a smartphone.
The value Qualcomm adds with AIFF lies in connecting neural models to Adreno's HPM, and packaging it as a complete product with a Vulkan-based SDK, plugins for major game engines, and development tools. The talk also showed an image quality comparison between AIFF and Snapdragon Game Super Resolution 2 using in-house demos. Both conditions involved upscaling from 960×540 to 1920×1080, but no numerical image quality metric was provided, nor was processing time disclosed. Power consumption and input latency also remain unknown. The comparison images were also selected by Qualcomm itself.
Snapdragon Summit 2026 will be held from September 22 to 24 on Maui, Hawaii, USA. Even if the correspondence with the next-generation SoC is revealed there, AIFF's real-world performance won't be determined until supported titles, extended operation on the same device, and input latency differences with generation disabled are all available. Whether the Pro-exclusive feature justifies the device price can only be judged—not by displayed frame count, but—once those three pieces of information are made public.