Nvidia RTX, DLSS and its incredible impact on future laptop gaming
If you’ve been at all tuning in with the progression of video games in recent years, you’ve probably come across the terms “RTX” or the fabled “Ray Tracing”.
For me, It’s hard to hear “ray tracing” without picturing shiny stormtrooper helmets.
What you’ll come to find out though, is that ray tracing technology is only one small part of “RTX”. And we’re here today to talk about what I think is the much more interesting aspect of Nvidia’s RTX technology: Deep Learning Super Sampling, or DLSS.
Part 1: Deep Learning
While the actual final effects of DLSS as a whole are calculated and processed in the “SS” part, it’s imperative that we know what deep learning is to understand how the incredible results are even possible.
Since “Machine Learning” and “Deep Learning” are entire sciences dedicated research teams at universities across the globe, and are considered the bleeding edge of computer science and informatics, don’t expect for us to get too technical. Broad strokes only today 🙂
As humanity progresses in its ever-terrifying pursuit after an artificial construct able to mimic a human’s ability to solve complex problems, we’ve reached a point where the next logical step in this path is creating an artificial neural network: A construct that functions similar to how a human brain would.
But what’s so special about the human brain that makes useful to us in so many industries, including video games?
The answer is of course its ability to learn. Hence, deep learning. “Deep”, in this context refers to the depth of the hidden layers as illustrated in the graphic above.
Makes sense, right? Now that we have a better understanding of that, we may ask the following: What exactly, in the context of video game graphics, is it learning to do?
Part 2: You Guessed it.. Super Sampling
Before we explore what Super Sampling means as it pertains to DLSS, I would first like to clarify that the term has existed long before Nvidia’s DLSS technology, to maybe help you avoid some confusion.
Ironically enough, the “old” meaning for Super Sampling (A.K.A SSAA, or Super Sampled Anti-Aliasing) is in a way an opposite of the image reconstruction process of DLSS, even though the actual desired function is, of course, a better looking final image.
SSAA is a form of Anti-Aliasing which is performed by rendering a desired object at a much higher resolution, then downscaling, or shrinking it to fit the target resolution of the game.
In direct contrast, Super Sampling as a part of Nvidia’s DLSS, allows the GPU to render the game at a much lower resolution. This lower resolution image then acts as the input of the Deep Learning model, like we mentioned before.
The input then goes through the DLSS process (the information is passed and processed in the hidden layers), and the resulting output is what the game WOULD look like if we played it at the original desired resolution. For Example:
Let’s say you wanna play Remedy’s excellent “Control” from 2019 in 4K with all the GPU-expensive ray tracing features (RTGI, RT Reflections, etc.) enabled. Without DLSS enabled, you’d be running the game the old fashioned way; your GPU is rendering every pixel in that 4K image, and so you can expect a modest framerate and overall performance even if you have a high-end graphics card.
The way this process would look with DLSS on (Specifically DLSS 2.0) is as follows:
Your GPU renders the game at 1080P , or around a quarter of the resolution of 4K.
With the image processing capabilities of DLSS 2.0, your 1080p frame is upscaled to as close to 4k as possible. In some cases, the final image may look better than the target resolution. In this case, native 4K.
As a result of the actual rendering happening in 1080p instead of native 4K, your graphics card (specifically the aforementioned “shader cores” section of the GPU) is rendering the image much more effortlessly, while leaving the upscaling to be done on the Tensor cores, which are specifically designed for that very purpose.
The result: framerates that simply would not be possible at the desired resolutions purely through raw power.
Sounds too good to be true? Well, right now it kind of is. This is because you need both an RTX graphics card AND a game that supports DLSS in order to reap these benefits.
Control was such a great example because the game has recently been updated with an “ultra performance” mode using the recently released DLSS 2.1 (currently only available for the RTX 3090 GPU).
This allows you to play the game at a whopping 8K (why…) while actually rendering at an internal resolution of 1440p.
If your mind is not blown at this point, then I’ve done a terrible job of explaining things 🙂
It’s okay, it took me a second to wrap my head around it all too!
Part 3: DLSS and Gaming Laptops
Picture this: you’ve been playing control and one day an update is released that allows you to suddenly play the game at a significantly improved framerate and maybe even at a higher resolution.
Better gameplay at no extra cost. Not just better gameplay, but potentially even the BEST gameplay you’ve experienced on your device.
With Nvidia constantly releasing newer versions of DLSS, each more impressive than the last, if you have a compatible RTX card, your setup is going to be playing every compatible game better over time. DLSS will improve itself through deep learning, and whenever Nvidia sees a stable enough upgrade to be released to the public, they will.
I remember watching a presentation by Elon Musk speaking about his electric automotive company Tesla, which employs similar Machine Learning techniques in order to gradually (and shockingly fast at that) improve the capabilities of its AutoPilot technology, which allows the car to drive itself on highways and even through city streets.
Elon Musk claims that because of the inevitable improvement the car’s systems will see as a result of the employment of ML techniques in combination with the head start Tesla has over other car manufacturers, “It would be financial suicide to buy anything other than a Tesla.”
Sure, Musk does have a knack for the dramatic at times, and it may be a bit of an exaggeration, but he does have a point, and it applies to Nvidia’s DLSS-capable RTX cards as well.
This is why I’ve recently criticized Razer for not including an RTX card in the newest model of the Razer Blade Stealth 13, where I said:
[So] while any RTX laptop, even one with a 2060 Max-Q for example, a graphics card not THAT much more powerful than the Stealth’s [GTX graphics card], will run any game with DLSS support (of which many, many more are currently being developed) at a constantly improving rate as the DLSS model matures and evolves, the [Blade’s GTX Card’s] capabilities will always be.. just that.
In other words: Buying a non-RTX card in a world with DLSS is, in a way, financial suicide.
What did we learn today, kids?
- Machine Learning is OP.
- AMD is very far behind and no-one is talking about it. (Remember Elon’s quote)
- If you’re gonna buy a graphics card for gaming, it would be insane not to go with an RTX GPU, because:
- Cyberpunk 2077 releases with DLSS support. A few months later Nvidia Control Panel says you have a driver update, and now the game will look AND run significantly better completely for free.
We live in exciting times. Thanks for reading, friends.