Nvidia’s RTX real-time ray-tracing
I’d like to do some additional experimentation with Nvidia’s RTX ray-tracing, but, even more than Radeon Anti-Lag, I don’t really have a satisfactory set-up for RTX testing. The only RTX game I have right now is Quake II RTX, and while I’m rather taken with Nvidia’s modification of the classic Id Software title, it operates in a radically different fashion from other RTX games.
Where Battlefield V uses RTX to simulate realistic reflections and Shadow of the Tomb Raider uses RTX for lifelike lighting and shadows, Quake II RTX relies on RTX entirely to light and shade the scenes. Nevermind that it mostly uses assets from 1997. I do own Chinese indie title Bright Memory (which was recently announced to be getting RTX support in an upcoming patch) as well as Remedy’s Control (releasing August 27), so perhaps I can investigate RTX performance some other time.
Even still, I have a fair bit to say on the topic of RTX. Regardless of what screeching fanboys will insist, ray-tracing in some form is almost certainly the next big step in real-time computer graphics. Rendering professionals always knew ray-tracing offered superior quality compared to rasterization. It simply wasn’t computationally feasible to use for games—at least, until now.
Of course, some folks say that it still isn’t computationally feasible. Contemporary implementations of RTX cast rather meager amounts of rays that bounce relatively few times, compared to offline ray-tracers. RTX games also make heavy use of an AI-powered de-noising filter to smooth out the grainy results. Despite all that, enabling RTX still has an enormous performance penalty in every game so far.
Personally, I think the end product looks great, more often than not. The question of whether those end results are worth kneecapping your game’s performance is a matter of taste, but judging from the buzz around the web, it seems like a significant portion of gamers aren’t exactly enthusiastic about the idea. The concept of lopping off half your framerate for improved image quality does seem to run contrary to the trend toward high-refresh monitors.
Hardware-savvy persons that I respect have drawn parallels between GeForce RTX and the technologies introduced by the old GeForce FX and GeForce 3 GPUs. The idea behind the comparison is that current GeForce cards’ relatively-poor RTX performance (compared to non-RTX performance) is essentially a matter of “growing pains” as we move into a new, hybrid rendering paradigm. In this comparison, the implication is that the next RTX-capable architecture will offer dramatic gains in terms of ray-tracing performance.
I hope that’s the case. However, R/T cores don’t perform AI inferencing. If Nvidia is to remain a “one-architecture company,” it’s difficult to imagine the next generation of GeForce cards offering the kind of leap in performance we saw between the GeForce 3 and the GeForce 4 Ti 4600, or between the GeForce FX 5800 and GeForce 6800 Ultra. Instead, it seems more likely that we will see another modest gain in RTX performance while significant portions of the chips’ die area are devoted to more tensor cores.
There’s also AMD’s position to consider; currently, Nvidia’s only real competitor in the PC graphics card space has no ray-tracing-specific hardware despite the existence of a vendor-agnostic accelerated ray-tracing standard in Microsoft’s DXR. I’ve talked about this before, in front-page news posts, but both Sony and Microsoft have made mention of ray-tracing in hype sessions for their next-generation consoles. AMD is, of course, supplying the hardware for both machines.
That doesn’t necessarily mean anything; as demonstrated by Crytek’s impressive Neon Noir demo above, ray-tracing can be performed (after a fashion) on modern graphics hardware without specialized hardware acceleration. Given that both Sony and Microsoft have also specifically mentioned “Navi” in reference to their upcoming hardware, and considering that “Navi” as we know it does not have dedicated ray-tracing hardware, that may be the route that the console creators are taking—or it may simply be buzzword bluster.
A brief look at power consumption and efficiency
Modern mid-range and high-end graphics cards are so much more power-thirsty than the CPUs they support that the classic PC expansion slot form factor has become inconvenient and even a bit ridiculous. An in-depth discussion of that topic will have to wait for another day, though, because I’ve got some graphics cards here that need their juice consumption examined.
To test each card’s idle power draw, I rebooted the machine after installing the card’s drivers, and then waited for 5 minutes to let Windows finish its various startup tasks. You’ll notice there’s no chart of idle power numbers here. I’m not completely confident in my numbers—I had some problems with Windows 10 background tasks—but by my measure, every card was within 15 watts of my 85-W baseline. It seems idle power draw is more or less a solved problem for both AMD and Nvidia.
Afterward, I loaded up Monster Hunter World in 3840×2160—so as to generate the maximum possible GPU load—and loaded into my test area in the Research Base area of the game. I watched the numbers flip by on my Kill-a-Watt for 30 seconds, recording the minimum and maximum values I saw, then averaging the two to get the numbers above.
There’s nothing particularly mind-blowing in the results; the Radeon RX 5700 XT is ever-so-slightly faster than the GeForce RTX 2060 Super, and it draws a bit more power. The GeForce RTX 2070 Super and RTX 2080 Super are based on the same chip, so their similarity seems simply-explained, and the big Pascal part with its 352-bit memory bus draws the most power.
As I mentioned way back on the testing methods page, the Radeon RX 580 and GeForce GTX 1660 Ti cards were loaners from friends, and I unfortunately didn’t have the tools on hand to measure their power consumption at the time.
I went ahead and charted the power consumption of each card under the cards’ aggregate 99th-percentile frame times. It’s a somewhat awkward measure of efficiency, but it gives you a reasonable impression of where each card falls compared to the others.
The latest GeForce cards are quite efficient overall, as expected, but AMD’s showing here isn’t awful. For a while now, enthusiasts have been shouting from the rooftops that Radeon cards become remarkably efficient when undervolted. Perhaps the slightly lower clocks of the base-model RX 5700 help reduce its power draw beyond the cuts made to its cores. Whatever the reason, it’s a fair bit more efficient than its faster sibling.
Meanwhile, the RTX 2070 Super and RTX 2080 Super in comparison to the GeForce GTX 1080 Ti give a nice little demonstration of the advancements made in the refreshed Turing: you can stick with the same power budget and get improved performance, or keep the same performance and save a little power. Impressive improvements, considering not only all the extra capabilities of the Turing GPUs, but also that Nvidia is still using a previous-generation fabrication process.