DLSS 4.5 arrived today with the latest Nvidia driver, and enthusiasts have already put the new upscaling model through its paces, benchmarking it on older RTX 30-series hardware. Mostly Positive Reviews on X, ran several benchmarks on their RTX 3080 Ti in Cyberpunk 2077, comparing DLSS 4.5 to 4.0, and saw up to a 24% performance hit from the more computationally intensive DLSS 4.5 model.
At the RT Ultra preset, Cyberpunk 2077 saw a 24% reduction in FPS, going from 42 FPS to 32 FPS when switching to DLSS 4.5. At 1440p, the performance drop went down to 14%, going from 72 FPS to 61 FPS, and at 1440p with ray-tracing disabled, performance dropped by 20%, going down from 108 FPS to 86 FPS.
Resolution and graphics settings | Frame Rate | DLSS Model |
4k / RT Ultra | 32 FPS (avg) | DLSS 4.5 (Preset M) - Quality Mode |
4k / RT Ultra | 42 FPS (avg) | DLSS 4.0 (Preset K) - Quality Mode |
1440p / RT Ultra | 61 FPS (avg) | DLSS 4.5 (Preset M) - Quality Mode |
1440p / RT Ultra | 72 FPS (avg) | DLSS 4.0 (Preset K) - Quality Mode |
1440p / Ultra (no RT) | 86 FPS (avg) | DLSS 4.5 (Preset M) - Quality Mode |
1440p / Ultra (no RT) | 108 FPS (avg) | DLSS 4.0 (Preset K) - Quality Mode |
These benchmark results are not surprising to see. Nvidia warned that it could not guarantee speedy performance from DLSS 4.5 on older RTX cards, specifically 30- and 20-series products DLSS 4.5 is a whopping 5x more computationally intensive on the tensor cores, but Nvidia is leveraging FP8 precision in RTX 40 and RTX 50 series cards for DLSS 4.5 to lessen the performance impact on newer cards.
Nvidia rep Jacob Freeman claims (on X) that DLSS 4.5 only runs 2-3% slower compared to DLSS 4.0 on RTX 50 series hardware. Sadly, RTX 30- and 20-series tensor cores lack FP8 capability.
In fact, owners of older RTX 30 and 20 series GPUs have been seeing regressions in DLSS computational performance since DLSS 4.0 launched. RTX 20, 30, and even RTX 40 series hardware run DLSS 4.0 upscaling slower compared to RTX 50 series GPUs. But it seems with DLSS 4.5, this effect has become substantially worse on older cards.
Confusing language from Nvidia - Nvidia's DLSS documentation contradicts its own blog posts
Negative performance isn't the only issue affecting DLSS 4.5; several Reddit threads have surfaced since Nvidia added the ability to override DLSS 4.5 in games with the Nvidia app, leading to confusion on how DLSS 4.5 overrides should be applied.
If you are unfamiliar with the inner workings of DLSS, DLSS takes advantage of .dll files and preset modes. The DLL files get updated with the latest model that Nvidia adds to DLSS, and normally, new preset modes get added to trigger the new model. For example, to override games with DLSS 4.5, you have to enable either the M preset or the L preset.
However, the DLSS 4.5 models may not be the best for use in all situations. Nvidia's latest DLSS programming guide for game developers suggests that the older DLSS 4.0 presets J and K should be the default model for DLAA, Quality, and Balanced modes, while DLSS 4.5 preset M is the default for Performance mode. According to Tom's Hardware's conversations with Nvidia on the ground at CES, the L preset is best used with the Ultra Performance mode.
You can, of course, use the DLSS 4.5 M model with higher-quality input scaling settings, but the computational expense of doing so may outweigh the image quality benefits of the new model.
If you want to override games with DLSS 4.5, it's worth doing your own research and testing all the presets to see which one looks the best to you for a given resolution and DLSS scaling factor.
Follow Tom's Hardware on Google News, or add us as a preferred source, to get our latest news, analysis, & reviews in your feeds.
