Nvidia DLSS 5 Runs on Apple Silicon at 2 FPS

Two independent developers have demonstrated that Nvidia's DLSS 5 neural rendering pipeline can run on Apple Silicon, a significant technical achievement that bypasses the company's proprietary hardware ecosystem. The feat, first reported by GN technics/hardware, relies on a chain of unofficial tools that bridge Apple's Metal graphics framework with Windows games. While this proves the underlying models are portable, the current result is a proof of concept rather than a usable product, with performance metrics that make it impractical for actual gaming.

The project involves two distinct engineering tasks stitched together to create a functional workaround. One developer ported the neural models to run natively on Apple's GPU architecture, while the other built a bridge using ReShade and the Game Porting Toolkit to connect this backend to Windows titles. This allows a game's rendering pipeline to hand off frames to a DLSS 5 model executing on Apple's silicon instead of an Nvidia RTX card, a complex interop that goes far beyond simple image upscaling.

Neural Rendering Requires Deep Integration

DLSS 5 is not a simple filter applied to the final image; it is a rendering stage that sits inside the game engine. It must interact with existing draw calls, motion vectors, and frame history in real-time. Reproducing this behavior outside of Nvidia's CUDA ecosystem and driver stack is a materially harder engineering task than porting a standalone image processor. The fact that it runs at all on Apple's tile-based architecture suggests the models are highly portable, provided the runtime is manually translated to fit the new hardware's constraints.

Performance Falls Far Short of Usability

The reported metrics explain why this is not yet shippable. One account cites a latency of 240 milliseconds, while another notes frame rates dropping to 2 frames per second during the DLSS pass. For context, competitive gaming typically targets total system latency under 30 milliseconds. A 2 FPS frame rate means the technology is technically functional but practically unusable for interactive media. This gap highlights the trade-off: the neural models are portable, but the current implementation lacks the optimized hardware acceleration needed to run at playable speeds.

Laptop Hardware Limits the Demonstration

The experiment reportedly used Apple's M5 Pro chip, a mid-tier processor aimed at prosumer laptops. This choice suggests the developers were working with hardware they already had on hand, rather than chasing best-case benchmarks on Apple's most powerful Max or Ultra silicon. While getting any version of DLSS 5 to execute on a laptop-class GPU is a nontrivial feat, it also means the current performance numbers likely understate the potential of the workaround on larger chips with higher memory bandwidth and core counts.

Ultimately, this development serves as a stress test for the portability of generative neural rendering. It confirms that the models are not locked to Nvidia's tensor cores, but the 2 FPS result serves as a clear reminder that hardware optimization is just as critical as the model architecture itself. Until the latency and frame rate issues are resolved, this remains a technical curiosity rather than a viable alternative for Mac gamers.