So just how much of a game-changer is DLSS? Well, based on the two demos we've seen, it's very exciting indeed. At the most fundamental level, a lower resolution image is generated, and then a deep learning algorithm programmatically upscales the image based on 'learnt behaviour' sourced by Nvidia's 'Saturn V' super-computer, studying extreme resolution game imagery, finding patterns in their composition and building it all into a reconstruction algorithm just a few megabytes in size. The bottom line is this: is the actual image being generated is of a lower resolution, it stands to reason that shading power required decreases, resulting in higher frame-rates.
But the idea of a super-computer being able to give instructions to a consumer level GPU to 'magic up' missing detail sounded rather implausible when it was first presented at Gamescom. However, based on the results I've seen so far, I'm happy to eat my words. We'll be talking more about image quality in due course, but suffice to say that we're getting a look here that exceeds the quality of checkerboarding and can even improve a native presentation (depending mostly on the quality of its anti-aliasing). But just how much of a performance gain do you get from it?
And so begins the benchmarking, where to begin with, we take a look at Epic's Infiltrator running at 4K resolution across GTX 1080, GTX 1080 Ti, RTX 2080 and RTX 2080 Ti. On the new cards, we're able to use DLSS as well as the standard TAA, while the older Pascal flagships are limited to standard TAA only. We're using our new benchmarking widget here, so desktop users can play the YouTube video to see real-time profiling, and can swap in and out any of the available profiles. On the barchart, click to swap between frame-rates and percentage differentials and mouse-over the GPU of your choice for the comparisons you're interested in. Lowest and highest scores here are frame-time averages, converted into frame-rate for easy reading. Mobile users can't run all of this tech efficiently, so there's a simpler table for those devices.
Epic Infiltrator Demo: TAA vs DLSS
The data is clear enough here and gives potential RTX 2080 buyers in particular food for thought. The GTX 1080 Ti remains an outstanding performer, it has more memory than the RTX 2080 and it's likely to get plenty of tempting deals in the short term. The RTX 2080 is almost 12 per cent faster, but as the other benchmarks shall reveal, it's not always such a clear-cut lead. However, with DLSS active, that 12 per cent lead rises to 58 per cent.
That's on par with the kind of performance boosts we saw between nine series and ten series cards and places a clear divide between Turing and Pascal products. Even without DLSS, the RTX 2080 offers a 41 per cent boost over its notional predecessor, the GTX 1080 - and this increases to a full doubling of performance once the reconstruction technique is enabled. Fascinating stuff. Also intriguing is that running with DLSS enabled allows the RTX 2080 to beat off the challenge of the RTX 2080 Ti running without DLSS enabled - in this case, there's a 13.5 per cent lead.
But obviously the RTX 2080 Ti is equipped with the same technology, and up against the GTX 1080 Ti it's 39 per cent faster with the like-for-like TAA enabled, rising to an impressive 95 per cent boost once DLSS is factored into the equation. In our general gameplay testing, we found that the RTX 2080 Ti is a phenomenal 4K performer just relying on standard rasterisation performance - but DLSS takes things into a whole different category. There are further 'in the moment' performance advantages beyond the average frame-rate boost we'll discuss in an upcoming feature, but suffice to say, the Epic demo here does a great job in selling DLSS. Image quality is generally a wash here - some scenes look better with standard TAA, others look less impressive than DLSS, but in motion, aside from variation in how specular shimmer resolves, it's really difficult to tell them apart. From our perspective, DLSS would be a lock where this an actual game... but it isn't. However, we do have a demo based on actual gaming imagery to use for verifying our results - Final Fantasy 15.
Final Fantasy 15 Demo: TAA vs DLSS
It's fair to say that Final Fantasy 15 is a great game, but its standalone benchmarking tool is remarkably poor, with often terrible performance - especially with Nvidia GameWorks enabled. We were given a version of the demo with settings locked seemingly at max (yes, with GameWorks active) and performance was OK, but marred by obvious stuttering. For our benchmark, we capture three stages of the bench where the stuttering is kept to a minimum. The idea here isn't to test Final Fantasy 15 with these GPUs, but rather to judge relative performance with the same workloads from areas that run fairly smoothly.
RTX 2080 performance with standard TAA reveals that the card enjoys a straight 30 per cent lead over GTX 1080, and it's basically on par with the GTX 1080 Ti - a state of affairs that's fairly common in the standard benchmarks to come. DLSS grants the RTX 2080 a further 39.5 per cent of raw performance, which clearly takes it well beyond the capabilities of even the most powerful Pascal cards. With DLSS active, the RTX 2080 offers an 81 per cent boost over GTX 1080. And once again we see that the RTX 2080 with DLSS enabled outperforms the 2080 Ti running on standard TAA. With RTX 2080 results this good, the impact on RTX 2080 Ti is even more profound. Again, with DLSS active, it's capable of delivering 80 per cent more performance than the GTX 1080 Ti, which does not have access to this technology.
With 28 titles confirmed for DLSS, it's clear that there is momentum behind adopting the technique. It requires a game engine that already uses temporal anti-aliasing and according to Nvidia, integration is fairly simple. We'll be deep-diving into image quality more intensively soon, but the fact is that Final Fantasy 15's TAA is actually rather blurry and not very impressive at all, and has genuine issues dealing with hair textures and other transparencies.
We'll be posting screenshot comparisons in due course, but myself and colleagues John Linneman and Alex Battaglia are agreed that in the Final Fantasy 15 demo at least, DLSS is not only providing these big improvements to performance, but it's delivering more detail and fewer artefacts than the game's standard TAA. It's not quite so clear-cut with Epic's Infiltrator demo - UE4 features one of the best temporal anti-aliasing solutions around - but the fact that it's so competitive is testament to DLSS's quality.
Personally, I think that DLSS is going to be essential in making ray tracing more performant in the short term. I've embedded a video here of Nvidia's Star Wars Reflection demo running on RTX 2080 and 2080 Ti. Bear in mind that up until the release of Turing, it took four Volta cards running in parallel to render this at 1080p24. Here, we're running the demo unlocked at 1440p with DLSS on both cards. The RTX 2080 Ti has the processing power and the memory to operate it at 2160p with DLSS - and although frame-rates are a lot lower, it is doing so consistently above the 24fps threshold. This gives you some idea not only of how ray tracing capabilities scale between RTX 2080 and 2080 Ti, but also how instrumental DLSS could be in getting frame-rates and/or resolutions higher when RT features are enabled.
One further note: because DLSS is based on machine learning, future driver updates will feature more refined versions of the technology, meaning that DLSS quality should improve generally over time.
GeForce RTX 2080/ RTX 2080 Ti Analysis
- Introduction, Hardware Breakdown, Turing Architecture features
- DLSS - Deep Learning Super-Sampling: Performance Analysis [This Page]
- Assassin's Creed Unity, Battlefield 1, Crysis 3, Far Cry 5 - Rasterisation Analysis Part 1
- Ghost Recon Wildlands, Rise of the Tomb Raider, Shadow of the Tomb Raider, The Witcher 3, Wolfenstein 2 - Rasterisation Analysis Part 2
- GeForce RTX 2080/ RTX 2080 Ti - the Digital Foundry verdict