Qualcomm’s Next Adreno GPU Adds AI Matrix Cores for Neural Fusion

Qualcomm has outlined a major change for the Adreno GPU in its next flagship Snapdragon platform: AI processing intended for graphics will move directly into the graphics pipeline. The company calls the feature set Adreno Neural Fusion, which combines neural processing, AI super resolution and frame generation in a single rendering pipeline. 2

The practical goal is familiar to PC gamers: render efficiently, then use AI to help produce a sharper, smoother-looking result. Qualcomm’s approach is not a universal graphics toggle, however. It depends on support and integration from game developers and engines.

Three Matrix Cores are built into the GPU

The new GPU includes three Adreno Matrix Cores, Qualcomm’s first AI-dedicated cores within the Adreno graphics pipeline. They are placed across three GPU slices, rather than the six-slice configuration suggested by earlier rumors. 27

Instead of sending graphics-related AI work out to a separate accelerator and back again, Qualcomm says the Matrix Cores can run AI models alongside rendering. The company’s case for this design is reduced data movement, lower latency and lower power overhead for AI-enhanced graphics tasks. 24

This does not replace Qualcomm’s Hexagon NPU. Hexagon remains the company’s dedicated on-device AI processor, while the Matrix Cores target AI work that benefits from staying close to graphics data and the rendering pipeline. 1923

What Adreno Neural Fusion is meant to improve

Neural Fusion brings together three components:

• neural processing
• AI super resolution
• AI frame generation 2

Qualcomm says the result should be higher visual quality at lower rendering cost and power use. Its stated image-quality benefits include sharper retained detail, fewer artifacts such as shimmering and ghosting, and more stable frame rates than earlier approaches. 214

Those are Qualcomm claims, not independently verified performance results. Actual image quality, latency and frame-rate gains will depend on the game, the implementation, a phone’s cooling and power limits, and the selected graphics settings.

Qualcomm claims up to 40% lower power for AI rendering

Qualcomm says Neural Fusion can use up to 40% less power than its previous solution for the relevant AI-rendering workloads. 214

The company attributes that potential saving to keeping AI models and graphics data local to the GPU instead of repeatedly moving them between subsystems. The 40% figure should be treated as a vendor-provided, workload-specific claim until third-party testing compares supported games on shipping devices.

Some reports have also cited an approximately 12% broader GPU power-efficiency improvement. Qualcomm’s Neural Fusion announcement, however, centers on the up-to-40% claim for its AI-rendering approach; a general 12% gain has not been established in the primary announcement provided here. 239

1.45GHz GPU slices and 18MB of Adreno HPM

Qualcomm says each of the three GPU slices operates at up to 1.45GHz. The design also retains 18MB of Adreno High Performance Memory (HPM), a low-latency memory resource directly available to the graphics subsystem. 710

HPM is intended to keep tile-based rendering data, frame buffers and compute workloads close to the GPU and the Matrix Cores. That local storage is central to Qualcomm’s latency and efficiency argument: less traffic to other parts of the system means AI-assisted rendering can remain in the graphics workflow. 243

Support in Unity and Unreal Engine matters

Qualcomm says Adreno Neural Fusion is supported across Unity and Unreal Engine. 2

That support is important because upscaling and frame generation need game-aware implementation to work well. In practice, availability will be determined by individual studios and titles, not solely by whether a phone contains the new Snapdragon hardware.

Where the GPU announcement fits in Qualcomm’s flagship preview

The Adreno disclosure is part of Qualcomm’s staggered preview of its next flagship Snapdragon platform. In August, Qualcomm announced a next-generation Oryon CPU that it says reaches 5GHz, along with Flex Cache, a dynamically allocated shared-cache architecture for heterogeneous cores. 21

The GPU reveal now fills in Qualcomm’s graphics and AI-rendering strategy ahead of Snapdragon Summit, September 22–24, 2026. 28 Qualcomm has not yet provided complete public specifications for the next platform’s Hexagon NPU in this GPU announcement, so its final AI performance and the division of work between Hexagon and the new Matrix Cores remain to be detailed.

For mobile gaming, the meaningful test will be whether supported titles can convert the new hardware into visibly cleaner upscaling, convincing frame generation and sustained performance without excessive battery cost. Qualcomm has described the architecture; shipping phones and independent testing will determine how much of that promise reaches players.