Modders have managed to run NVIDIA’s DLSS 5 neural rendering technology on Apple Silicon, showing that the underlying neural network is not limited entirely to GeForce hardware.
The achievement follows similar community experiments involving AMD Radeon and Intel Arc GPUs. However, the Apple implementation is still far from practical for gaming because its processing latency is much higher than on NVIDIA hardware.
Testing on an M5 Pro MacBook Pro showed a neural rendering pass taking around 240 milliseconds at 1440p. For comparison, an RTX 5050 reportedly completes a similar workload in roughly 24 milliseconds.
That puts the current Apple Silicon implementation at about ten times the inference latency.
Apple implementation uses MLX and Metal
The experimental work is based on two community projects called MLX DLSS and DLSSMac.
MLX DLSS runs the neural rendering model through Apple’s MLX machine learning framework and Metal graphics API. DLSSMac then connects that backend to Windows games using ReShade and Apple’s Game Porting Toolkit.
Neither project has official support from NVIDIA or Apple.
The implementation also requires neural model data extracted from NVIDIA’s DLSS libraries, meaning this is currently an experimental community effort rather than a consumer ready feature.
Apple Silicon cannot use NVIDIA Tensor Cores, which are specifically designed to accelerate the low precision AI workloads used by DLSS. Instead, the neural workload has to be executed through Apple’s available GPU and machine learning resources.
Additional overhead comes from ReShade and the Game Porting Toolkit translation layer.
Performance remains the biggest problem
Current results show that compatibility is possible, but performance is still a major limitation.
| Hardware | Test result | Resolution |
|---|---|---|
| Apple M5 Pro | Around 240 ms neural pass | 1440p |
| Apple M5 Pro test build | Around 21.96 FPS | 960 × 600 |
| RTX 5050 | Around 24 ms neural pass | Comparable test |
| Radeon RX 9070 XT | Around 33 FPS | 1080p |
| Intel Arc 140V | Around 3 to 5 FPS | 720p |
The M5 Pro test reportedly reached an estimated 21.96 FPS, but only at 960 × 600 resolution.
That is enough to demonstrate that the system works, but it is not yet close to the performance needed for a normal gaming experience.
AMD and Intel have also run DLSS 5
Apple Silicon is not the first non NVIDIA platform to run the technology.
A separate Radeon project has recreated the required neural runtime through HIP and reportedly reaches around 33 FPS at 1080p on an RX 9070 XT.
Another implementation runs the same 71 block neural network on Intel Arc hardware using XMX acceleration and Vulkan cooperative matrices.
That version reportedly manages around 3 to 5 FPS at 720p.
These experiments suggest that DLSS 5 neural rendering is portable at the software level once developers reproduce enough of the required runtime.
The larger challenge is processing the workload efficiently.
NVIDIA hardware still has a major advantage
DLSS 5 is officially designed for NVIDIA’s GeForce RTX 50 Series.
Blackwell GPUs include Tensor Cores built specifically for AI calculations and support native FP4 processing. That gives NVIDIA hardware a substantial efficiency advantage when executing the low precision neural workloads used by the technology.
Apple, AMD and Intel hardware can perform similar calculations, but current unofficial implementations do not have the same optimized software stack or dedicated execution path.
That helps explain why the performance gap remains so large despite successful demonstrations.
Neural rendering could become more portable over time
These projects are important because they separate software compatibility from practical performance.
Running the model on Radeon, Arc and Apple Silicon demonstrates that the neural network itself is not fundamentally tied to one GPU architecture.
That does not mean DLSS 5 will become an officially supported cross platform technology.
NVIDIA still targets the feature at its own RTX hardware, while AMD, Intel and Apple have not announced plans to support NVIDIA’s implementation.
The experiments instead show what developers may eventually achieve with their own neural rendering technologies.
For now, Apple Silicon can technically run DLSS 5 neural rendering, but current performance remains too slow for practical use in demanding games. Further optimization would be needed before this type of unofficial implementation could become useful beyond technical demonstrations.
Discussion (0)
Be the first to comment.