NVIDIA's latest AI-powered upscaling technology, DLSS 4.5, has begun rolling out to users, promising a significant leap in image clarity and detail reconstruction. However, early testing reveals a substantial trade-off, particularly for owners of older RTX 20 and 30 series graphics cards. While the new model produces visibly crisper results, it comes with a heavy computational burden that can slash frame rates by 20% or more on hardware lacking specific architectural features. This development highlights the growing divergence in AI performance between GPU generations and forces users to make a critical choice between visual fidelity and smooth performance.
The Core of DLSS 4.5: A More Demanding AI Model
At its heart, DLSS 4.5 introduces a second-generation Transformer-based AI model, a more complex and computationally intensive successor to the model used in DLSS 4.0. This new neural network is designed to perform more sophisticated analysis and reconstruction of game frames, leading to the reported improvements in fine detail, texture sharpness, and overall image stability. NVIDIA has confirmed that this new model is approximately five times more compute-intensive than its predecessor. This raw increase in processing demand is the primary reason for the performance impact observed across the board, though its severity is not evenly distributed.
A Generational Divide: The Critical Role of FP8 Precision
The performance disparity between GPU families stems from a key hardware difference: support for FP8 (8-bit floating-point) precision in tensor core operations. NVIDIA's current-generation RTX 40 series and the newly announced RTX 50 series natively support FP8, which allows their tensor cores to handle the massive calculations of DLSS 4.5 much more efficiently. An NVIDIA representative, Jacob Freeman, stated that on RTX 50 series hardware, the performance penalty for enabling DLSS 4.5 is a mere 2-3% compared to DLSS 4.0. In contrast, older RTX 20 and 30 series tensor cores lack FP8 support, forcing them to run the new model using less efficient data formats, which drastically increases processing time and VRAM requirements.
Quantifying the Performance Hit on Older Hardware
Community testing, which began shortly after the driver release on January 6, 2026, has provided concrete numbers on the performance regression. Benchmarks conducted on an RTX 3080 Ti in Cyberpunk 2077 at 4K resolution with Ray Tracing Ultra settings showed a severe 24% drop in average frame rate, falling from over 40 FPS with DLSS 4.0 to around 32 FPS with DLSS 4.5. At 1440p, the hit was measured at 14% with ray tracing enabled and 20% with it disabled. These findings were corroborated by other users, including one with an RTX 4060 laptop GPU, who observed a nearly 16% performance decrease. The trend is clear: the older the architecture, the heavier the toll for the improved image quality.
Reported Performance Impact of DLSS 4.5 vs. DLSS 4.0 (RTX 3080 Ti)
| Game & Settings | DLSS 4.0 Avg FPS | DLSS 4.5 Avg FPS | Performance Drop |
|---|---|---|---|
| Cyberpunk 2077, 4K, RT Ultra | 42 | 32 | 24% |
| Cyberpunk 2077, 1440p, RT Ultra | 72 | 61 | 14% |
| Cyberpunk 2077, 1440p, No RT | 108 | 86 | 20% |
| The Last of Us Part 2, 1440p High | 154 | 135 | 14% |
The Compounding Issue of Increased VRAM Consumption
Beyond raw compute performance, DLSS 4.5 also places a greater strain on video memory. According to NVIDIA's documentation and user reports, the new model requires significantly more VRAM. For RTX 40 and 50 series cards, the increase is between 40-53%. However, for RTX 20 and 30 series cards, the VRAM requirement skyrockets by 87-103%, essentially doubling. This creates a potential double-whammy for cards with limited memory buffers, such as 8GB models. These GPUs may not only suffer from the computational slowdown but could also encounter stuttering or crashes due to VRAM exhaustion, especially at higher resolutions or with texture-heavy games.
Reported VRAM Usage Increase with DLSS 4.5
- RTX 40 & 50 Series: 40% to 53% increase
- RTX 20 & 30 Series: 87% to 103% increase (approximately double)
Weighing the Visual Benefits Against the Performance Cost
The critical question for users is whether the visual upgrade justifies the performance loss. Early side-by-side comparisons, such as those shared from an RTX 4060 laptop, demonstrate that DLSS 4.5 can render noticeably sharper details in elements like foliage, character models, and environmental textures. The improvement is tangible. For gamers playing older or less demanding titles where they already enjoy high frame rates (e.g., 100-200 FPS), sacrificing 15-20% of that performance for a cleaner image may be a worthwhile trade-off. Conversely, for those playing cutting-edge, graphically intensive games where every frame counts, the performance penalty may make DLSS 4.5 impractical on older hardware.
Navigating Implementation and User Confusion
The rollout of DLSS 4.5 has introduced some complexity for end-users. NVIDIA has enabled model overriding through its app, allowing users to force the new DLSS 4.5 model (identified by presets like 'M' and 'L') into games. However, the company's own guidance suggests that the newer, more intensive models are not always the optimal choice. Their developer documentation indicates that the older DLSS 4.0 presets ('J' and 'K') should remain the default for DLAA, Quality, and Balanced modes, with the DLSS 4.5 'M' preset targeted primarily for Performance mode. This has led to confusion, underscoring the need for users to experiment with different presets to find the best balance of image quality and performance for their specific hardware and game settings.
Conclusion: A Step Forward with Clear Caveats
DLSS 4.5 represents a genuine advancement in AI-driven image reconstruction, pushing the boundaries of visual fidelity in real-time rendering. However, its arrival solidifies a growing reality in GPU technology: cutting-edge AI features are increasingly optimized for the latest hardware architectures. For owners of RTX 40 and 50 series cards, DLSS 4.5 is a nearly free upgrade offering superior image quality. For those with RTX 20 and 30 series cards, it presents a difficult compromise. The technology is supported, but its use requires a careful, case-by-case evaluation of whether the crisper visuals are worth the sometimes-substantial frame rate cost and increased system resource demands.