I wouldn't use PUBG as a benchmark for anything, but GTA, Assassin's Creed, the Witcher, Fallout or lretty much any major AAA open-world game runs at 30fps. Considering I would hope for world sizes and densitities to increase significantly, cpu is going to need a significant boost. Just hoping for enough to hit 60 in all titles, not just the linear games. I'd prefer 1440p60 to 4k30 if given the option.
Could next gen consoles focus mainly on CPU?
- Thread starter Theeoo
- Start date
It's definitely doable even 4K60. There are games today that do that on X1X that are very much AAA looking titles.
The real question is whether we need _more_ CPU on the silicon size of things. If a change in architecture is already sufficient, I don't see the need to scale larger even more.
I'd like to see next gen start in the 18+TF range to really make a generational gap between this gen and the next. But I can see that as occupying a large portion of the silicon. Which is also why I feel if there are other technologies that can relieve the CPU (and there are many and it continues to increase), or we can find ways to offload more to the GPU (and developers are doing more of this), then this becomes achievable because we don't need to dramatically increase the CPU size to make 60fps or more nuanced AI for instance.
The real question is whether we need _more_ CPU on the silicon size of things. If a change in architecture is already sufficient, I don't see the need to scale larger even more.
I'd like to see next gen start in the 18+TF range to really make a generational gap between this gen and the next. But I can see that as occupying a large portion of the silicon. Which is also why I feel if there are other technologies that can relieve the CPU (and there are many and it continues to increase), or we can find ways to offload more to the GPU (and developers are doing more of this), then this becomes achievable because we don't need to dramatically increase the CPU size to make 60fps or more nuanced AI for instance.
I've yet to encounter any game doing game AI with neural nets. I haven't Googled any examples or papers either. Game AI is decision making based on evaluations, with every entity having a different situation to evaluate such that there's little parallelism that can be exploited, AFAICS. There's definitely scope for aiding AI through compute, through the spatial modelling and whatnot I've suggested, but I don't think multiple AI agents is a good match for GPU unless there's a completely new paradigm I've missed. Which no-one's written a paper about yet?
You could also use computer learning to study the player and come up with different strategies, but putting them into practice with agent decision making?
I think PUBG is a great example of how the right hardware enabling the right software enables innovation. And also of how the games people are most excited to actually play may be ones that place the concept ahead of the optimisation.
I think eight Zen 2 cores - two modules - running at about 3 gHz would be a good compromise. More threads for developers who have the bandwdith to thread further, but even for those who don't still something like a 3x increase in CPU performance over PS4Bone. And that immediately puts significant gains in anyone's hands for either high frame rate versions of conceptually risky 20 fps chuggers or massive scope increases for current 30 and 60 fps showcases.
That still leaves you with the vast majority of your 300~350 mm^2 SoC for GPU and a big fat GDDR6 interface.
I'm sure you could. But as you allude, this would still need to feed back into behaviours that games designers are honing to create fun experience, with enemies and characters that fit certain gameplay or story needs.
And the vast majority of game designers - even the really good ones - aren't writing papers on neural nets or low level GPU programming. The more abstract the tools they can use, and the more human friendly, the better.
I can't honestly imagine any game devs even exploring it! They're already maxing the GPU with pretties and have reliable AI mechanisms they can control as needed. Bethesda had to dumb down their AI because it could break the game, as an example of what can go wrong with highly flexible systems. So who's spending significant time trying to make GPUs do what the CPU is doing and one of the CPU's main jobs if it isn't going to be redundant? It's not even going to be creating draw calls from now on. 
I've just identified an area where computer learning may well be applicable.
Currently I'm wrestling with an efficient, effective AI that'll evade objects while pursuing targets in a top-down 2D shooter. It's a problem where the only tool I have is lots of ray casts. I keep thinking about alternatives if only I had better data - if only I could perform a space test and get how much coverage is there, or get the distance to the edge of the sprite on one side versus the other. And as you look at the problem, you appreciate humans just look at the thing and know where to go, not just based on the object's placements but also their motion.
And that there is where a computer can use modern Big Data AI, I think - have it look at the game, looking at the same screen the player sees, and perform actions based on derived learning. Or rather, I guess, render a specific simplified view for when they aren't in the same space or context that the computer just look at and decide what to do. But all the decision making can be based on the same Big Data processing the players do.
Of course, realistically that's not going to work for any complex game. GTA VI won't be able to have a virtual brain for every resident, drawing the world from their POV and analysing it. We have entire custom GPGPU processors to do just that for a single car! But it does show to me at least one aspect where GPU acceleration can be applied to game AI.
Currently I'm wrestling with an efficient, effective AI that'll evade objects while pursuing targets in a top-down 2D shooter. It's a problem where the only tool I have is lots of ray casts. I keep thinking about alternatives if only I had better data - if only I could perform a space test and get how much coverage is there, or get the distance to the edge of the sprite on one side versus the other. And as you look at the problem, you appreciate humans just look at the thing and know where to go, not just based on the object's placements but also their motion.
And that there is where a computer can use modern Big Data AI, I think - have it look at the game, looking at the same screen the player sees, and perform actions based on derived learning. Or rather, I guess, render a specific simplified view for when they aren't in the same space or context that the computer just look at and decide what to do. But all the decision making can be based on the same Big Data processing the players do.
Of course, realistically that's not going to work for any complex game. GTA VI won't be able to have a virtual brain for every resident, drawing the world from their POV and analysing it. We have entire custom GPGPU processors to do just that for a single car! But it does show to me at least one aspect where GPU acceleration can be applied to game AI.
D
Deleted member 13524
Guest
Why should there be papers talking about using A.I. neural network inference specifically in games? What is it about a game engine that makes it hard to implement a neural network that turns outputs based on a set of inputs after training?
Regardless, GPGPU crowd simulation A.I. (path-finding with several inputs each with different weight factors.. pretty much a NN if you ask me..) was shown over 9 years ago on DX10 hardware, with results claiming "5-39x faster than CPU":
http://s08.idav.ucdavis.edu/shopf-crowd-simulation-in-froblins.pdf
And this is pre-OpenCL, pre-DX11 Compute Shader and most of all, pre-APIs with low overhead.
So why was it that we needed super-CPUs for next-gen consoles?
Regardless, GPGPU crowd simulation A.I. (path-finding with several inputs each with different weight factors.. pretty much a NN if you ask me..) was shown over 9 years ago on DX10 hardware, with results claiming "5-39x faster than CPU":
http://s08.idav.ucdavis.edu/shopf-crowd-simulation-in-froblins.pdf
And this is pre-OpenCL, pre-DX11 Compute Shader and most of all, pre-APIs with low overhead.
So why was it that we needed super-CPUs for next-gen consoles?
You sure are ambitious. I believe the next console generation will top out at 10 / 12 TF. Why? Because desktop GPU's performance upgrades in single precision TFlops have been slowing down, not accelerating. If PS5 / NextBox is coming in 2 years as the rumours go, no way they will be at 18TF. That will most likely be on the ballpark, give or take 4 or 5 TF of what a high end desktop GPU will be able to do and consoles are never as powerful.
PS4 had 40% of the TF of the GTX780Ti. For a similar ratio and 18TF, you would have to have desktop GPU capable of 60TF!
Sure PS4Pro and Xbox One X have disrupted this trend, but they came 2 years+ after the original releases.
Last edited:
Do you want your game to be fun to play? That requires - in almost all cases - a lot of explicit rule tweaking by someone who understands games, after interpreting extremely subjective feedback.
There are papers talking about game application of all kinds of technologies. Why would this be any different?
Supplementing traditional game development with additional technologies is great, and exciting, but how do you propose to replace current game design and development processes with this, and how do you propose to do it within the next two to three years?
Look at all those interesting behaviours! No wonder gamers have been chomping at the bit to play this for 9 years.
Because reality.
(And a 3gHz Ryzen equivalent is not really a super CPU now, and it will decidedly mainstream on 7nm).
This may be more appropriate an example to give context for a shipped game's needs/requirements:
https://www.gdcvault.com/play/1022411/Massive-Crowd-on-Assassin-s
Whether that's translatable to the GPU environment (or if they'd budget it)... don't know.
hopeful and I definitely agree it would be highly improbable. But it's also the right number to aim for. I'm personally looking at 4 years from now to next gen (2022), not two years mind you.When silicon stops moving, it stops moving for everyone. So I would assume console technology will eventually catch up with PC. 4 years is a long time. PC GPUS move to 7nm this year. 5nm in 3 years time. By the 4th year the process and a lot of the funding is going through maturity and bringing down the cost to fab since we can't shrink anymore. There would be money put elsewhere I imagine.
PC is at 11.5 TF today. 7nm say 22 TF. 5nm say 30TF.
In 4 years we have 4x more power from base model to mid-gen model. In 4 more years assuming exponential scaling, it was be 16x of the base model. So 16-18 TF vs PC 30TF?
Those are my pie in the sky numbers.
Consoles will never catch up with discrete graphics. The bulk of console sales happens at the sub $300 price point. Once silicon scaling dies, that'll be the average price point. About 60% of that is going into the APU+RAM, ~$180. A discrete GPU don't have bundled controllers, storage devices, PSU so you end up with a higher fraction of the $300 going towards silicon in a GPU.
And $300 is just the starting point for discrete graphics (in the future).
There is a much bigger chance of PC integrated graphics catching up with consoles. The main difference is in the memory subsystem. If (when) high bandwidth technologies (Intel's EMIB/AMD's interposers) mature and costs come down, Intel and AMD can add much more silicon to their existing APUs.
Cheers
D
Deleted member 11852
Guest
Sometimes oft seemingly complex problems have simple solutions. Have a read of the 'AI' used for the for Inky, Clyde, Blinky and Pinky in Pac Man, because combining some of traits may work. Most tracking algorithms of this nature have been discovered, published and ripe for the picking and ironically they're decades old and simple because they had to work low powered hardware.
The biggest challenge in games in that the amount of computational resources available per frame is disproportionately small compared the the processing budget dedicated to rendering. Think about what realtime games have to do compared to turn-based games like Civ VI which can take 30 seconds per turn even on a meaty PC. It's also on iPad and as much as I'm curious I'm not willing to spend $40 to find out how late game turns take even on an 2017 iPad Pro.
as always, your posts have been insightful.
Battlefield 4. Multiplayer game. Stable 60 fps on both base consoles. Looks way better than PUBG on base Xbox One. PUBG supports up to 100 players (Battlefield 4 up to 64). PUBG is more, but not dramatically so.
Battlefield 4 frame rate tests (solid 60 fps):
Battlefield 4 player count (up to 64) patch discussion with devs:
https://www.vg247.com/2013/12/04/ba...rs-online-took-priority-over-1080p-says-dice/
Very good looking 60 fps online game with 64+ players is certainly doable with the base 1.6 GHz Jaguar CPU. But it requires lots and lots of work and technical talent. Ryzen would make console development easier for sure.
Games tend to have different AI requirements to jobs where most of the research is done. All these AI developments do one thing - learn to read a language or recognise an image or find oil reserves from data crunching - and not select from multiple alternative actions with different outcomes.
You could implement a neural net, but what exactly is it learning and how does it implement it?
You seem to be extrapolating from a vague sense of what's accomplished. This demo isn't AI in effect - this is a fluid simulation with a bunch of particles following with it. Now take the idea of thousands of AI agents, but given different priorities. What if some are looking for each other? Some are trying to decide whether to carry on or go home? At the moment every one of them is following the 'flow' of the movement, but what if one needs to be able to determine whether to follow everyone else or veer off to the top of the mountain, going against the flow? Or decide whether it's better to intercept that agent moving at that velocity or hang back to intercept that other agent, or go to a different location entirely? This demo has many iterations of the same simple behaviour, exactly the sort of parallel workload GPUs are good at. Games need massively diverse behaviours, potentially huge numbers of decisions per agent, and which can be controlled according to the designers. Bethesda's Radiant AI ended up breaking quests for the player so they needed to control it.
I'm not going to say its impossible to use compute to produce effective game AI, but it's absolutely not trivial. Anyone wanting to try that will be undertaking a huge amount of work working out the models and implementations that'll allow it. A workload similar to the change from single threaded game engines to multithreaded, which devs only undertook because CPUs stalled on the single threaded performance progression and had to become multicore. AI on the CPU is a known quantity that can be tweaked and controlled without needing to reinvent the wheel. What's probably needed if there's to be a paradigm shift is a bunch of academics creating a number of papers of game AI simulation (complex decision making based on lots of interrelated parameters). I think it worth noting that this demo is 9 years old. How many more advanced GPU AI demo's have there been since then? If devs are sitting on this gold-mine of AI power that'll speed their games up, why aren't they using it?
I imagine one of many problems that could arise is that if you put AI on the GPU, it's only going to be for console, that's one aspect. The other is that if the results don't come back fast enough, you've stalled the whole pipeline.
I think we could see the AI go onto the GPU if there is available and reliable amount GPU power guaranteed in a certain time frame.
Battlefield 4 is certainly an impressive game (all the more so for its age), but online mode is a long way from being a stable 60 fps. It's often in the 50s, sometimes in the 40s, and can crater into the 30~40 fps range. Again, impressive given what's it's doing and how old it is, but it's a long way from a stable 60hz.
I can't know for sure whether the CPU is implemented in these drops, but given where some of them happen and what's going on onscreen (in terms of explosions, effects etc) I think it has to be a factor in some.
I think it's possible that a game with similar player counts and more objects and types of interaction could see increasingly stressful hotspots. Players have a nasty habit of not always distributing themselves evenly and predictably.
A CPU 3 ~ 4 times faster should be able to run something like B4 at a solid 60hz and with increased complexity and/or player counts to boot.
And it'd open up ambitiously scaled game making to far smaller and less well funded developers, increasing the opportunities for innovation.
djskribbles
Legend
A ~31% increase in clock speed does wonders in BF4. It's still not locked, but not too far off.Battlefield 4 is certainly an impressive game (all the more so for its age), but online mode is a long way from being a stable 60 fps. It's often in the 50s, sometimes in the 40s, and can crater into the 30~40 fps range. Again, impressive given what's it's doing and how old it is, but it's a long way from a stable 60hz.
I can't know for sure whether the CPU is implemented in these drops, but given where some of them happen and what's going on onscreen (in terms of explosions, effects etc) I think it has to be a factor in some.
I think it's possible that a game with similar player counts and more objects and types of interaction could see increasingly stressful hotspots. Players have a nasty habit of not always distributing themselves evenly and predictably.
A CPU 3 ~ 4 times faster should be able to run something like B4 at a solid 60hz and with increased complexity and/or player counts to boot.
And it'd open up ambitiously scaled game making to far smaller and less well funded developers, increasing the opportunities for innovation.
