The FUTURE of graphics

unless you've been living under a rock you might have noticed how Graphics in games seem to be stagnating despite the impressive ui5 demos I've been showcasing here on the channel like on the video I released a few days ago which you should go watch in Practical terms the biggest games of the year like Starfield cyber Punk Phantom Liberty lies of PE Hogwarts Legacy Assassin's Creed Mirage or even the ui5 based Lords of the Fallen don't look any better than say Red Dead Redemption 2 which came out 5 years ago Hardware to run games is also stagnating with small incremental uplifts and generations spread further apart well necessity is the mother of all invention and it seems we are at a pivotal point when it comes to games graphics with a true Revolution forming on the horizon one that might turn the hardware industry on its head before we continue a quick word from our sponsor today private internet access if like me you're someone who relies on public Wi-Fi like the kind you get in cafes or shopping malls or airports you are exposing your sensitive data over a completely unencrypted Network which could put you at risk as hackers that are connected to that same network can steal your personal data with ease that's when a VPN can be incredibly effective as it hides your data by running it through their worldclass server infrastructure so that you can rest assured that your browser activity is not being seen by anyone else Pia have a no log policy that there were adamant to defend in court and have been independently audited to show that they don't store any data about you it's fantastic if you do work on the go as I do but it's also super useful if you like to stream content from different regions I'm a fan of snobby art housee Indie movies and many of the Publishers don't let me access their services because I'm based in Portugal even though there aren't any alternatives for watching those highbrow films where you can contemplate a tree for 5 minutes straight Pia solves that problem as you can change your IP address to one of 84 countries and all of the 50 US states Pia is available on Windows Macos Android Linux iOS and many more and you only need one Pia subscription to protect all the devices you have at the same time so your entire household your workspace and all your devices around the world all protected with just one subscription sign up with a link in the description of this video to get an 83% off discount on private internet access so that's just $2.3 per month and you also get four extra months completely free a huge thanks to private internet access for sponsoring today's video the graphics pipeline hasn't changed much in the last couple of decades there are more polygons on screen more effects higher resolution textures photogrammetry has added to the realism and real-time rate racing has added more Sheen but at the core Graphics are still just polygons rendered in the viewport textures apply to them lighting and shadows to make the whole thing more believable unre Engine 5 has advanced graphics with Lumen and especially Nite which allows for a greater level of detail but it's still the same old polygon with textures lighting and shadows just more of it with more stuff to render AMD and Nvidia and now Intel have released ever more complex gpus with more transistors and more fixed function on ever more expensive transistor nodes in 2017 the flagship Enthusiast class GPU at the time the 1080ti would set you back $700 a mere 6 years later and the flagship Enthusiast GPU costs today around $1,800 an almost 160% increas in cost in order to drive Graphics that like I said in the intro don't really look that much better than games that that were around when the 1080ti was King something's got to change if Graphics are too advanc and let's face it great graphics still sell games I've been talking about neurographics and that's the direction that I think things are going what's interesting is that it seems changes are happening so fast that it's difficult to predict who exactly will come out on top most would think Nvidia has the advantage given their dominance in AI but I'm not so sure one thing's for sure Nvidia is that at the Forefront when it comes to innovation in AI right now not just in Hardware but particularly in software one example of this is Masterpiece x a new tax2 3D AI playground created by Masterpiece studio in collaboration with Nvidia Masterpiece studio is aiming to bring 3d creation to as many people as possible without years of learning complicated software like blender or Maya or 3D Studio Max using generative Ai and Nvidia casso GPU optimize inference the creation process here happens through an API as a service workflow rather than on a local machine running expensive hardware and complicated software all you need is a browser and you don't need any prior knowledge of 3D modeling just prompt to program and you get a 3D model as with any Tex to anything software the more detailed The Prompt the better the result as it stands and as you can see in these examples the model is not quite ready for High Fidelity AAA assets just yet but it's good enough for prototyping and the resulting asset is compatible with 3D software like blender so you can tweak it if you have the knowledge for it so no Hardware is required no software is required and no skills are required you just type in what you want to create and a 3D model that's compatible with game engines is produced the potential for disruption here is tremendous this is the next Frontier and a tapping Stone to a true generative pipeline line back in 2018 Nvidia demoed an interactive completely AI generated 3D world that world was created based on training of real world driving footage Nvidia extracted the high level semantics so trees buildings and other vehicles lamp posts Etc attributed different colors to different objects and then the network converts these representations into complete images using Unreal Engine 4 as a framework at the time looking fairly rudimentary since then with huge training models we no longer need footage we just need taxs to describe what we want it's a Quantum Leap in just a few years and as far as implementation goes we already have dlss 3 generating fake frames in between real frames with frame generation so gradually of course this won't happen overnight we will see more and more neural rendering being added to the traditional rendering pipeline first with assets creation on the part of artists but further along with the end engine itself or the graphics API or the driver creating visual representations of things that only exist in the engine either as basic elements or even as a text description so for instance taking Starfield as an example instead of there being a planet surface that has a bunch of 3D models of rocks that were placed there either manually or procedurally by Bethesda there could just be a yellow Cube and the graphics driver tells a portion of the hardware that whenever that yellow Cube appears it should generate a realis 3D model of a rock or not even a 3D model but a neural model a new type of asset that is only a proxy since the generative process would be handled by a portion of the hardware specifically designed for this task like a dpu or an npu this could be done much faster than rendering a traditional 3D object now imagine an engine where you don't even need a yellow Cube but just a tax description and the game engine is no longer a framework for polygons in texture and lighting but rather a framework to get generated content that is consistent with a game Creator's Vision so that the game feels coherent a more fine grained approach would be that a studio would only hire concept artists and from the concept designs this hypothetical engine would generate infinite variations of assets and infinite game worlds that were consistent with the Studio's Vision based on the concepts created by the artists obviously this can only go so far and there will have to be curation of content scripted events there's no Magic Bullet but it would massively speed up both the creation part for the game designers and the rendering part for the users if this sounds too out there well all of what I described already exists in rudimentary forms and in some cases like with Masterpiece X in a fairly Advanced and usable form neuro textures were also announced by Nvidia so instead of loading high resolution textures you can set the game to neural textures which will load low resolution textures but then generate them with more detail therefore requiring less vram and less horsepower to achieve good performance with convincing results a new technology that I briefly mentioned in my last video is 3D gausian splatting in short g-s Splats similar to Nerfs 3D Garian also fall into the category of Radiance fields and are a new way of creating a 3D scene using only a video or a sequence of images the great thing about g- plats is that they are much faster to train and you get photo realistic results in a matter of minutes or an hour depending on the video you are using in simple terms using a mathematical method called structure from motion 3D points are calculated based on the differences between the images and someone already managed to import a gplat into an real engine for fortnite using Niagara particles it's clearly early days but just like NV is turning lowr textures into neural textures for much higher quality a similar method could be used to turn gats into usable photorealistic environments with more convincing results than even photogrammetry now guess what you don't need anymore in this looming fake world of generated Graphics assuming this new neural pipeline is indeed feasible you no longer need a GPU what you need is an AI accelerator and a GPU might not necessarily be the right tool for that or a GPU might end up being relegated to image output and processing but not rendering this open UPS the hardware Market to new players because even though Nvidia has a hard start in such a pivotal moment being stuck in the old ways as Nvidia is might actually be an obstacle companies like T torren come to mind as well as Qualcomm who have been investing heavily in AI fixed function for a while now qualcomm's recently announced Orion CPU is dubbed in their marketing material as built for generative Ai and their collaboration with Microsoft seems to already bear fruits as the screenshot of task manager on a system running the elite X so shows a dedicated Snapdragon npu for AI compute tasks there will be interesting to see how the elite AE performs versus an Nvidia GPU in image generation I think if a killer app emerges like good quality video generation enabled by an integrated npu Nvidia could have a problem on their hands in the near term with implications for the long term if the industry adopts something else other than gpus for this new era of AI content generation and rendering AMD also included AI accelerators in their radon 7000 gpus and you can now run stable diffusion automatic 1111 on those accelerators to generate an image from text in 3 seconds Intel has something similar cooking with open Veno on the arc gpus and they even have a web UI that you can use to tweak the results so for now while gpus are the go-to tool for AI no one will make a Dy in Nvidia dominance but at the pace things are progressing out of nowhere we could see a new paradigm for graphics in games emerge as a consequence of such a pivot generate 3D assets from taxed or generate a video from taxed and then turn that video into a gplat and import it into a game engine that's already possible imagine in a couple of years where we'll be for years I've been talking about the threat of A6 or application specific devices to Nvidia strategy remember that while in Nvidia gpus have massive throughput and the throughput we need for the edge they are nowhere near the Energy Efficiency required I've talked about tasla as an example a company that ditched Nvidia when they realized Nvidia would be way over the power budget that they could fit in a Tesla vehicle and over time that protection of mine has been materializing with A6 maintaining a leading efficiency and now starting to get very close to Nvidia gpus in terms of performance the blue dots here are fpga by the way and you can see they are terrible at efficiency they're good for prototyping but not for deployment this chart doesn't tell the whole truth of course as this is Peak Performance and assumes full utilization and architecture is only as good as you can keep it filled you might remember how amd's Vega ended up being bottlenecked by poor utilization why does this happen well because the data to be worked on is rarely perfectly symmetrical whatever the workload type how do you solve this you solve this with heterogeneous systems which is the route that AMD is taking with their apus although in a very limited form for now I think what Apple has been doing is probably a better example the ultimate solution however and if Graphics in games do indeed transition to a neural rendering pipeline would be something more akin to this this is the Tesla FSD it features three clusters with quad a72 arm processors two large npus a GPU and a bunch of smaller devices for security video encoding Etc but obviously for the desktop stop it would not come in this form factor but rather disaggregate it so each of these elements being its own chiplet what's interesting about this possible future is that new hardware makers could focus on specific chiplets to add to a common chiplet library and I think the GPU as we know it could take a back seat in this future having this much flexibility with help with occupancy resulting much better efficiency and reduce costs that's where I see things headed for the future of Graphics from a Hardware perspective and don't be surprised if the next version of Windows requires a dpu for the optimal experience with features being gimped out for those who don't have the dedicated Hardware in the meantime there's a lot in the traditional rendering pipeline that can be accelerated Nvidia has already introduced neuro Radiance caching for instance for real-time path tracing dlss is here to stay even though I'm personally not a fan and acid creation is going through a massive revolution of course it wouldn't be a CX video if I didn't talk about olad displays and that's the shest way of upgrading the quality of your graphics right now it sounds like Alienware and others are about to introduce high resolution High refresh rate oleds in early 2024 so start saving up for those I might be reviewing something related to that soon so be sure to subscribe to the channel also be sure to check out my video from a few days ago showcasing some of the coolest looking ui5 demos out currently thanks for watching and until the next one [Music] f

Channel: Coreteks

Views: 35,725

Keywords: nvidia, graphics, future of graphics, future gpus, 4080ti, 4090, 5090, rtx 4080, rtx 4080ti, rtx 4080 super, artificial intelligence, generative graphics, generative ai, amd, intel, neural textures, masterpiece x, next-gen graphics

Id: GKnr27-_6Mw

Length: 15min 46sec (946 seconds)

Published: Sun Oct 29 2023