admin
The recent emergence of NVIDIA’s DLSS 5 has sparked intense scrutiny within the hardware benchmarking community, particularly regarding its implementation of neural rendering technology. As the latest iteration of Deep Learning Super Sampling (DLSS), version 5 represents a fundamental shift from traditional temporal upscaling toward a more comprehensive neural-based image synthesis. While early reports indicated that DLSS 5 could place a significant strain on contemporary PC hardware, new data emerging from the developer community suggests that the technology is particularly punishing for non-native architectures. Specifically, recent tests conducted on Apple’s M5 Pro silicon reveal a stark disparity in performance when compared to dedicated NVIDIA hardware, highlighting a critical architectural gap in Apple’s latest high-performance chips.
The Experimental Framework: Porting DLSS 5 to Apple Silicon
The testing of DLSS 5 on Apple hardware was not an official endeavor by NVIDIA or Apple, but rather a complex technical feat achieved by independent developers. The process involved a multi-layered software stack designed to bridge the gap between NVIDIA’s proprietary libraries and Apple’s Metal graphics framework. Developer @iamwavecut spearheaded the effort by porting the neural rendering models of DLSS 5 to run natively on Apple Silicon using Metal. This was further supplemented by the work of developer @Mappsnet7, who utilized a combination of ReShade and Apple’s Game Porting Toolkit (GPTK) 4.0b2 to hook the Metal backend into Windows-based titles.
The primary testbed for this experiment was the 16-core GPU variant of the Apple M5 Pro. The objective was to determine if the "Neural Rendering" capabilities of DLSS 5—which promise lifelike visuals by replacing traditional rasterized lighting and shading with AI-generated approximations—could be effectively handled by Apple’s Unified Memory Architecture and Neural Engine.
Visual Fidelity vs. Computational Reality
Upon successful injection into the rendering pipeline, the visual results on the M5 Pro were described as transformative. The implementation of DLSS 5 brought about a significant evolution in image quality, specifically in the realms of surface lighting, ambient occlusion, and depth realism. The neural models effectively mitigated many of the artifacts common in traditional screen-space reflections and shadow mapping, producing a scene that appeared significantly more grounded and photorealistic.
However, these aesthetic gains were offset by a catastrophic decline in performance. While the image quality reached new heights, the 16-core GPU of the M5 Pro struggled to maintain a functional frame rate. Benchmarks recorded during the test showed the M5 Pro delivering a mere 2.32 frames per second (FPS) at a resolution of 1440p. Perhaps more detrimental to the user experience was the impact on input latency. The tests revealed that input latency ballooned to 240ms—nearly ten times the latency observed on an entry-level desktop RTX 5050. This level of performance renders games effectively unplayable, transforming a real-time interactive experience into a slow-motion slideshow.
The Technical Bottleneck: FP8 vs. FP16
The primary reason for the M5 Pro’s lackluster performance in this scenario lies in the specific mathematical requirements of the DLSS 5 neural models. According to analysis provided by the hardware community, specifically Redditor "AnyPomelo3352," DLSS 5 is heavily optimized for FP8 (8-bit floating point) precision. FP8 has become the industry standard for high-speed AI inference because it allows for significantly higher throughput and lower power consumption than higher-precision formats like FP16 or FP32, provided the hardware has dedicated units to handle it.
NVIDIA’s latest architectures, including the Blackwell-based RTX 50-series, feature dedicated FP8 hardware units designed specifically to accelerate these types of neural rendering workloads. The desktop RTX 5050, for instance, is capable of delivering approximately 105 TOPS (Tera Operations Per Second) of FP8 performance. In contrast, the Apple M5 Pro lacks dedicated FP8 hardware acceleration within its GPU cores.
Because the M5 Pro does not support native FP8 processing, it is forced to route the DLSS 5 neural rendering workload through its FP16 (16-bit floating point) path. While FP16 offers higher precision, it is computationally more expensive and offers no performance benefit for models designed for FP8. The M5 Pro’s 16-core GPU maxes out at roughly 26 TOPS of FP8-equivalent performance when forced through this sub-optimal path. This represents a nearly 400% performance deficit compared to NVIDIA’s budget-tier desktop hardware, explaining the massive disparity in frame rates and latency.
The Role of the Game Porting Toolkit and Software Overhead
It is also essential to consider the role of the software environment in these benchmarks. Running Windows-based games on macOS via the Game Porting Toolkit 4.0b2 introduces an inherent layer of translation overhead. GPTK works by translating DirectX 12 calls into Metal commands in real-time. When an additional layer like ReShade is used to hook neural inference models into this pipeline, the cumulative overhead becomes substantial.
While Apple has made significant strides in reducing this overhead with each iteration of GPTK, the complexity of DLSS 5’s neural rendering—which requires constant communication between the GPU and the neural inference engine—exacerbates the inefficiencies of the translation layer. Even if the M5 Pro possessed superior raw compute power, the lack of a native, first-party implementation of such neural technologies would likely still result in a performance penalty compared to a native Windows/NVIDIA environment.
Chronology of DLSS Development and the Shift to Neural Rendering
To understand the significance of these findings, one must look at the evolution of NVIDIA’s upscaling technology. The trajectory of DLSS has moved steadily away from simple spatial upscaling toward complex AI-driven synthesis:
- DLSS 1.0 (2018): Introduced basic AI upscaling but suffered from "smearing" and ghosting artifacts.
- DLSS 2.0 (2020): Shifted to temporal upscaling, using motion vectors to significantly improve stability and image quality.
- DLSS 3.0 (2022): Introduced Frame Generation, using AI to insert entirely new frames between rendered ones, though it required the Ada Lovelace architecture.
- DLSS 3.5 (2023): Introduced Ray Reconstruction, replacing hand-tuned denoisers with a neural network to improve ray-traced image quality.
- DLSS 5 (Current): Moves toward "Full Neural Rendering," where the AI is no longer just "fixing" or "scaling" an image but is actively participating in the generation of light, shadows, and textures.
Each step in this timeline has demanded more specialized hardware. Apple, while a leader in mobile and efficient computing, has focused its Neural Engine primarily on photography, voice recognition, and machine learning tasks within the macOS ecosystem, rather than the high-throughput, low-latency demands of real-time gaming graphics.
Broader Implications for Apple Silicon and the Gaming Industry
The results of the DLSS 5 testing on the M5 Pro serve as a wake-up call for Apple’s gaming ambitions. Over the past three years, Apple has aggressively marketed its silicon as a viable alternative for "AAA" gaming, bringing titles like Resident Evil Village, Death Stranding, and Assassin’s Creed Shadows to the Mac. However, the industry is clearly moving toward a future where neural rendering is not an optional "extra" but a core component of the graphics pipeline.
If future game engines are built around the assumption of FP8-accelerated neural rendering, Apple’s current hardware trajectory may find itself several generations behind. The M5 Pro and M5 Max, despite their impressive performance in video editing and general productivity, appear ill-equipped for the specific computational demands of the next era of graphics.
This raises significant questions regarding the upcoming Apple M6 chip. Industry analysts suggest that for Apple to remain competitive in the high-end graphics space, the M6 must incorporate dedicated FP8 hardware units and a more robust interface between the GPU and the Neural Engine. Without these architectural changes, Mac gaming may remain relegated to porting older titles rather than participating in the cutting edge of real-time visual technology.
Industry Reaction and Conclusion
While neither NVIDIA nor Apple has officially commented on these unofficial community benchmarks, the reaction from the developer community has been one of cautious observation. Some developers argue that the M5 Pro’s performance is not a fair reflection of the chip’s potential, given the layers of translation involved. Others, however, point to the TOPS (Tera Operations Per Second) data as an undeniable proof of a hardware-level bottleneck.
The experiment conducted by @iamwavecut and @Mappsnet7 demonstrates that while Apple Silicon is capable of producing stunning visuals through neural rendering, it currently lacks the specialized "horsepower" to do so at playable speeds. As NVIDIA continues to push the boundaries of what is possible with DLSS 5 and beyond, the pressure is on Apple to evolve its silicon architecture to include the low-precision, high-throughput hardware necessary for the future of interactive entertainment. For now, DLSS 5 remains a tantalizing glimpse into a photorealistic future—one that, for Mac users, remains just out of reach.
