We recently saw AMD executives David Wang and Rick Bergman discuss the power of AI and what they hope to achieve using it in AMD’s next generation of graphics cards, which will be built on the RDNA 4 architecture.
RDNA 4 will include AMD’s second-generation AI Accelerator cores - and while Team Red is definitely still behind Nvidia and the mighty AI-powered performance of the RTX 4090’s Tensor cores, it sounds like Wang and Bergman want to take AI in some very different directions to Nvidia.
While one of Nvidia’s big selling points right now is DLSS 3, a GPU-integrated program that uses deep-learning AI algorithms to ‘upscale’ gameplay to a higher resolution to provide better framerates at 1440p, 4K, and 8K, AMD hasn’t taken the same route. Team Red’s answer to DLSS is FidelityFX Super Resolution (FSR for short), which notably doesn’t use AI for its upscaling processes.
In fact, AMD made kind of a big deal about this last year, claiming that AI just isn’t necessary for this sort of upscaling tech. It was a cheeky shot at Nvidia (and Intel, whose Arc GPUs use deep-learning tech called XeSS, similar to DLSS), but it looks like AMD is now climbing aboard the AI bandwagon.
Hostile AI and not-so-hostile AI
Before we go any further, I’m going to make a quick distinction to - hopefully! - make this article a bit easier to follow. When I talk about ‘AI’ from this point onwards, I’m going to be specifically referring to the processes that are used to control non-player characters (NPCs) and objects in games. So ‘enemy AI’ would refer to the code that dictates the actions of an enemy unit in a game, like a hostile soldier or a boss monster.
Meanwhile, I’ll use ‘machine learning’ to refer to what many consider to be actual AI - that is to say, software capable of learning, adapting, and self-modifying based on new information, like popular chatbot ChatGPT, or the deep-learning processes used in Nvidia’s DLSS software.
With that clarification sorted, let’s examine what Wang and Bergman suggested. There are some very promising ideas regarding a dedicated machine learning pipeline for rendering processes which could hugely improve graphics performance, but I don’t want to talk about that - I want to talk about the ways in which AMD believes deep learning could improve the gameplay experience itself.
The power of AI in games
The example used by Wang discussed AI pathfinding in games - this is an area where there’s still plenty of room for improvement on the development side, so it’s a good place to start.
Pathfinding, for anyone who is unaware, is the way in which a character or object in a game navigates through the game world. It can be a tough nut to crack; anyone who has played an Elder Scrolls game can probably recall a time when a hostile bandit got stuck in a bush or ran in circles while trying to reach you. Machine learning processes could almost certainly do a better job of navigating a virtual space than pre-programmed scripts controlling movement, so Wang’s idea makes perfect sense.
I’d like to take things a little bit further, though. If machine learning can be used to dictate how an AI unit in a game moves, what else could it do? Could it control the way enemies and allies react to your actions? Could it write realistic unscripted dialogue on the fly?
Games that ‘learn’ from your decisions aren’t anything new. The excellent Alien: Isolation is a prime example, with the titular xenomorph that stalks you throughout the game also taking stock of how you choose to counter it - use the flamethrower on it one too many times, for example, and it’ll learn to keep its distanced when you have it on hand. Metal Gear Solid V: The Phantom Pain has a less sophisticated but no less clever system; if you rely too heavily on headshots or gas grenades, future missions will feature soldiers freshly equipped with helmets or gas masks.
But machine learning has the potential to take these ideas and dial them up to 11, all while also creating a more immersive, more responsive experience for the player. I’m picturing a WWII shooter where every individual trooper on the beaches of Normandy is able to react convincingly to everything going on around them, instead of just getting stuck in a loop of screaming and trying to flee when disarmed.
I’m starting to worry about this
There are some, shall we say, worrying moral implications of this. The idea of playing a shooter where every enemy soldier carries the pseudo-sentience of an advanced machine-learning program is a tad scary, especially once we consider the degrees of intelligence that computers could achieve in the future.
The ‘White Christmas’ episode of the tech-horror anthology Black Mirror explores the idea of using digital human clones as glorified smart-home software, and what is a digital human clone if not simply a sufficiently-advanced machine learning program? I’m really not sure how I feel about the idea of blowing up virtual combatants who have a degree of actual sentience - especially not when I can fail the level and restart, forcing all those AIs to go through hell all over again.
I feel like I’m writing the plot of a dystopian thriller and I’m also not really qualified to discuss the philosophical ramifications of all this, though, so I’ll talk about something I definitely am qualified to discuss: graphics cards. If games start to employ AMD’s Accelerator cores to improve the game experience, suddenly we could find ourselves in a situation where some games don’t just look better on a certain GPU - they are better.
This would, in my opinion, be an unmitigated disaster for the PC gaming space. Imagine being told that you can’t play the latest game - or that you’ll be getting a dumbed-down version of it - because you don’t have the latest best graphics card! AMD’s Radeon RX 7900 XTX is great, but it can’t make games play any differently to any other GPU right now. It’s something that could lead to the endless console exclusivity war bleeding into the PC arena, and I don’t think any PC gamer wants that.
Unfortunately, in the capitalist hellscape we currently live in, it sounds all too plausible. AMD - and, for that matter, Nvidia and Intel - would no doubt jump on that sort of control over the market. If AMD could offer a superior version of a game provided you play it on Team Red’s hardware, the money-printing potential is extremely tangible. So, uh… maybe don’t do that, guys?