Ray tracing now runs on older Nvidia GPUs - but how fast is it?

With the arrival of Nvidia's RTX video cards and support for real-time ray tracing in both the DirectX 12 and Vulkan graphics APIs, we're looking at one possible future for graphics technology - but it is one mired in controversy because up until this week, only RTX GPUs could run DXR-enabled software, meaning only a very limited number of PC users could access ray tracing functionality. But now, the floodgates are open: Nvidia's latest driver allows its 10 and 16-series GPUs to run DXR software too. But is any kind of playable experience possible on cards without RTX's hardware accelerated support?

We tested DXR software across three cards that lack Nvidia's RT hardware - specifically the GTX 1060 6GB (the least capable compatible card), the GTX 1660 Ti (Turing with RT cores) and the top-end gamer's 10-series offering, the excellent GTX 1080 Ti. To get a sense of comparison against the RTX line, we opted to bench the RTX 2080 and the RTX 2060. It's an interesting assortment of cards: the RTX 2080 typically runs non-RT workloads at a similar speed to GTX 1080 Ti, while the 2060 is the cheapest card in the RTX line.

But what we quickly discovered is that outside of specially prepared ray tracing workloads, gaming benchmarks flatter to deceive the older cards. What quickly became evident is that RT workloads introduce such an enormous variability into gameplay that canned benchmarks only have limited relevancy compared to an actual experience of playing the game, and sometimes there's basically no correlation at all. Ray tracing can add a baseline cost to a game, but in scenes heavy in RT components, a 10-series card crumples while a 20-series GPU powers through. As we discovered, what you test in-game needs to be very carefully selected.

First up though, it's worth explaining what sets an RTX card apart from a GTX product and by extension, how ray tracing works. First of all, the GPU creates the structure that the rays will be shot at - the Bounding Volume Hierarchy, or BVH. Then the rays are shot at the BVH - and that's what Nvidia's RT core accelerates. Finally, there's another expensive step: denoising. Because only a relatively tiny amount of rays can be shot, the results you get back are very patchy/noisy - this needs to be processed into something more visually palatable. The bottom line though? In three demanding processes, only one is accelerated by RTX - and the RT core can be emulated by compute shaders, which is how 10-series cards can be brought into the mix.

Pascal Ray Tracing Tested! GTX 1080 Ti vs RTX 2080/ RTX 2060 + More

And to be clear here, the entire ray tracing procedure is carried out by Windows via the DXR API. DXR doesn't require hardware acceleration, your GPU just needs to have sufficient RAM and DX12 support. While 10/16-series ray tracing support is now enabled, there is nothing to stop AMD releasing their own DXR support for its existing line-up of graphics hardware - and owing to their heavy bias towards GPU compute, we could see some surprises there.

But returning to the here and now, Nvidia talked to us about how different workloads run better on non-hardware accelerated GPUs. The ray traced shadows seen in Nixxes' port of Shadow of the Tomb Raider were seen as an example of a fairly light load, the implication being that GTX cards should compare reasonably well to RTX cards. The reflection tech seen in Battlefield 5 is harder to process without RTX but is still seen as viable, while advanced reflections, path-tracing etc could be beyond older hardware. From our perspective, the important thing to note is this: games use hybrid rendering, a combination of standard rasterisation and ray tracing which means that as more RT elements dominate the scene, RTX will deliver vastly improved results. But if RT isn't prevalant, a GTX 1080 Ti could - in theory - outperform an RTX 2060.

To get an idea of how important acceleration is, we'll kick off with 3DMark's Port Royal benchmark, which is designed to put RT workloads to the forefront. This shows why the RT core in Turing RTX cards is so important. While the RTX 2080 and GTX 1080 Ti duke it out on standard 3D gaming with similar results, in this benchmark, the new card delivers a vast 169.4 per cent improvement in performance. The much cheaper RTX 2060 also has a commanding lead over the 1080 Ti, to the tune of 70.2 per cent. The non-Turing GTX 1660 Ti struggles somewhat, while the GTX 1060 is a write-off.

Battlefield 5 is our next port of call, and it's here where the ambiguities in benchmarking and performance begin to make themselves known. We chose the beginning of the Provence stage with its engine-driven cutscene, followed by a stroll through the Nazi-infested woods. It's a challenging stage in its own right, and the ground has its fair share of big puddles boasting real-time reflections delivered by an ingenious combination of DXR and standard screen-space reflections.

Letting the benchmark plays out delivers what - on the face of it - is a remarkable result. Despite its hardware acceleration and a huge 70 per cent lead in Port Royal, the RTX 2060 is less than eight per cent faster than the GTX 1080 Ti. And it's true that you can lower DXR to medium and much of the campaign will run at 60 frames per second or higher, just like the RTX 2060. In fact, thanks to the Ti's 11 gigs of RAM, you don't have to reduce texture quality from ultra, either. This is a prime requisite for smooth BF5 RT gaming on the 2060.

The benchmark table here is quite remarkable, but what we need to stress is that it cannot be taken as a realistic comparison of RTX and GTX cards in a ray tracing workload, as we'll explain shortly.

So why does the benchmark say one thing while we say another? The bottom line is that overall performance across the run of play is nowhere near as consistent as the RTX 2060, which leads us to an inescapable conclusion - just like a great many of the benchmarks built into a great many games, our like-for-like test runs aren't indicative of the actual gameplay experience.

If you can put up with a good degree of variance in performance, the GTX 1080 Ti can produce a decent enough ray traced experience for Battlefield 5 at 1080p resolution. However, any kind of scene that is heavy on reflections - particularly when reflecting transparencies - sees non-RTX performance drastically reduce. Alex Battaglia came up with his own RT Doomsday scenario, the so-called puddle of doom. Look into this body of water and behold as we start to see a similar level of differentiation in Battlefield 5 as we did in Port Royal. And remember that we're testing RTX 2060 as the comparison point - a $350 GPU up against a much more expensive big-chip Pascal challenger. Suffice to say that the equivalent RTX 2080 blows it away in all things ray traced.

To be honest, publishing the benchmark above does still trouble us somewhat because it's not about frame-rate averages, it's about finding the right workloads to test, so while the bench is indicative of non-heavy RT workloads, here's a shot to illustrate just wide the differential can be, even with the lower-power RTX 2060. Check out the video above to see how close we could get to 60fps DXR gameplay with the entry-level GTX 1060 6GB - it works, but let's just say that it's far from optimal.

Next up, Shadow of the Tomb Raider - and a fascinating take from Nixxes and Nvidia on how ray tracing can be used for something different than the signature shiny bits. We'd also recommend this developer interview on the topic, as it's fascinating stuff. Nvidia reckons that shadows and ambient occlusion are another relatively light workload for RT cores, and on top of that, the implication of DXR varies radically in Shadow's quality modes.

Ultra is exactly that, but medium is something very different - and something we may expect to see in next-gen console games assuming they have no hardware-accelerated RT features. Think of DXR medium here as 'ultra+' - essentially augmenting shadowmaps with strategically placed RT shadows. That explains why the non-RT cards in our grouping perform so well here, with the GTX 1080 Ti even able to outpace RTX 2060 by a whopping 32 per cent. And even on ultra, it's ten per cent to the better.

Of course, this doesn't tell us the whole story. Even on medium settings, there are DXR-heavy scenes in gameplay can wipe off 100fps from the GTX 1080 Ti's performance level, while other scenes barely impact frame-rates at all and run smoothly at 1080p resolution even with DXR set to ultra. This is the DXR game perhaps best suited to playing on a non-RT card as good performance is possible with some tweaks, but there'll always be those cliff-edge moments where frame-rate can tank without hardware acceleration.

There's more information and benchmarks in the embedded video at the top of the page, but they only serve to confirm what we've already discovered in Battlefield 5 and Shadow of the Tomb Raider - these games were never built to be run on non-accelerated hardware and while benchmarks in a vacuum may impress, once you actually start to play for any length of time, fluctuations in performance are just too impactful. Our Metro Exodus benchmark numbers suggested that the 1080 Ti wasn't that much slower than RTX 2060, but the bottom line is that we could run the game at 1080p60 with RT on the 2060, while the GTX 1080 Ti struggled to maintain 1080p30 across the run of play.

So is this exercise spectacularly pointless then? Definitely not! First of all, the more hardware to support DXR the better. The lower the baseline support in term of GPU power, the more likely it is that developers will produce 'bang for the buck' quality modes, like the medium DXR setting in Shadow of the Tomb Raider. As a proving ground for next-gen console ray tracing implementations, this could be invaluable.

Secondly, Nvidia's RTX line has been hamstrung by GTX comparisons that show little progression from one generation to the next. By allowing DXR to run on Pascal, users can sample ray tracing and perhaps 'feel the difference' a potential Turing upgrade could deliver. Some might say that it's about about encouraging upgrades - so a tad cynical, perhaps - but AMD is almost certainly going to enable its own DXR support for existing GPUs at some point too, so it's good that owners of Nvidia legacy hardware (and non-RT Turing) can try out the feature too.