Could domain specific hardware accelerators take us to game engine specific hardware?

Apologies you didn't read my previous post I shall repeat a point I made earlier...

It does not have to be a dedicated card it can be a System On A Chip with the CPU, GPU, EPU, NPU and HBM3 memory all stacked together for optimal latency.

The chip industry is adopting 3D stacking and Chiplet layouts that will allow for more complex SOC chips that are tailored for their market.

 

So are you saying you want a special SOC to put on a dedicated card to use in addition to the existing CPU in a system, or are you referring to replacing the existing CPU with a SOC? I definitely expect to see more amazing SOC/APU solutions replacing existing CPUs, but I do not expect to see additional dedicated hardware cards being added to most consumer systems.

 

The article you linked to was about implementing particle systems purely on the GPU. That is exciting for particle systems, but it is a different use case than ECS. In that linked article, the position updates are being done on the GPU and then that data is left on the GPU. The CPU is not receiving any position data back from the GPU in that particle system implementation. That is fine for a particle system, because the main thread in a game normally does not need to care about specific particle locations at runtime

A use case for ECS might be a large RTS game with hundreds of thousands of units moving around and thousands of projectiles flying around. Think StarCraft but on a massive scale. The individual units in StarCraft would need more intelligence than a particle system could provide. That where something like ECS comes in. ECS can provide massive scalability and plenty of intelligence per unit. ECS is much more flexible than a particle system.

Moving ECS to a GPU would be different than moving a particle system to a GPU. Computing all of the unit position updates each frame could be offloaded to the GPU in both cases, but for ECS all of that data would need to get retrieved from the GPU each frame. The ECS position data would also need to be sent to the GPU each frame as well, because local input actions and network based actions would affect the unit positions. With ECS, you cannot just push the unit position data to the GPU once and let it run the same way you would with a particle system.

But like I said already, don't just link to somebody else's unrelated work. Implement a proof of concept and prove us all wrong.

 

All you would get is a performance bottleneck. Feel free to implement a proof of concept to prove me wrong.

Moving the ECS code onto a dedicated hardware device would mean your game's main thread would need to send all of the ECS data to the card, wait for the card to process the data set, and then retrieve that data set back from the card. All of that would need to happen each frame. The cost of sending and receiving that data every frame would quickly exceed any performance gains you might get from the dedicated hardware.

With a GPU based particle system, your CPU sends the work to the GPU one time and then lets the GPU handle the particle system completely async. That frees up the CPU. With your idea of offloading ECS to a dedicated card, you would tie up the CPU with a bunch of extra work sending and receiving data with the add-on card every frame. It would reduce the performance of your game.

 

Another thing to consider here is the idea of a Physics Processing Unit. A company called Ageia came up with the idea of a physics software library that could be used in games and would offload physics calculations to a dedicated hardware accelerator called a PhysX card. The cards offered a performance advantage compared to a purely CPU based implementation.

The Ageia PhysX cards were very rare. Very few games supported them, and very few consumers purchased the cards. The idea of a standard physics library designed for games was actually more valuable than the hardware cards themselves. Nvidia bought the tech, released PhysX to the industry, killed the PhysX cards, and then released PhysX support in their GPU drivers.

Even today, nearly nobody configures their PC to have a dedicated GPU just for PhysX processing. Even gamers with several GPUs never consider that.

 

You keep posting completely unrelated concepts.

 

It was not a benefit LOL
Consoles has always been inferior versions due to their inferior price and hardware compared to PC.

 

You just proved, how little you know about another topic, you try to discuss. You quote random stuff, without either reading, or understanding the content. As of example from the link, it talks about CPU vs GPU particles. It mentioned about 150k particles at given time on CPU at low frame rate 20 to approx 30fps (optimized). Yet we saw 100k+ of objects moving on the scene with ECS in Unity c#, at high fps rate, rather than directly writing for OpenGL app.

As I would love tech discussions, this again feels one sided.

As other suggested, I am now more convinced, OP don't understand ECS implementation in Unity.

 

It's 2019, enough with the console vs pc garbage

 

I was not refering any console wars (i have both hardware), i was just responding to to @Arowx post saying that it was a bad example refering consoles hardware.

Anyway, you know you don't need to comment everyone response with your own interpretation LOL

Have a nice day.

 

Your first line did that. Your next did not. Inferior is a subjective term in this instance because technical performance is not the key deciding factor in whether or not a game is good or even well received.

 

It's forums main issue, people interpret words with their own vision.

I was referring hardware only, why do you need to discuss game success ?
You seem to turn everything so complicated lol

 

Hardware inferiority is largely a myth.

 

You are a bot ? LOL

 

No, I'm a game developer who's worked on loads of different platforms on loads of wildly different hardware. It does not matter if an Xbox One or PS4 is basically just a bunch of low-spec or custom built components compared to a PC. It only just mattered that the Wii was performantly a kinda beefed up Gamecube. The platform advantages between PC, console, handheld, and even mobile almost completely even out when you think about them as more than spec sheets.

 

More detailed graphics, 4K , RTX , better and more physics, PC or console or mobile won't run the same games with the same quality. Hardware defines the graphics and physics quality you can get.
I think you are a bot that will find some random ideas to tell above statement is not true LOL

It was fun, have a good day.

 

Wherein I guess you forget that most games don't even remotely touch these features, especially in the indie space, but even in the AAA space because the games are optimized for consoles first I guess? Phones, consoles, PCs, and handhelds also all have remarkably different playstyles from one another.

 

My pc can't do 4k and rtx, that's not inherent to pc but to specs, and console have the avdntage of a single optimization target and architecture sugar that allow them to punch above their weight. For example console with fixed function shader vs programmable shader at the same spec, or historically the megadrive vs snes.

What we need is based spec to result ratio to compare, then console win.

@Arowx
ECS is really a baked methodology (as in syntax) to take adventage of a based architecture layout (memory cache). "Dsp chip" also have the probleme of communication bottleneck between compute core, even in soc, in fact that's kinda what a gpu is, a lot of small core and that's why they have trouble communicating together (ie why query another core slow things down), cpu core are more independant by increasing complexity and localising more work (hence why less of them).

If you were to build an engine on a chip you would map the various core based on the underlying engine assumption about data layout, access and compute need. So the question is not really to add another dsp core, but to create a dedicated layout, that's kinda what console are and they special chip for input,sound security etc... for example the ps4 has an adventage with two bus to allow access to the gpu to support better gpu compute support, so you can do stuff that are send to the cpu while rendering other stuff on the cpu. But then when you want to support all sorts of game and application you tend to support generic architecture, and console don't fall too far from a basic pc architecture aside the sugar.

In fact it would make more sense to have a game on a chip (ie a specific architecture) than an engine on a chip, though more specific engine could have an in between. But then specific architecture call for specific language too, because the language is really just the interface to the underlying architecture (hence shader language to talk to the gpu), which is another pain, the more specific you go the less you can reuse common knowledge and tools, that was teh pain of old console till the at least the ps2 era. Which bring an intermediary step that is micro code instead of fpga, ie write your own low level instruction (like in the n64).

 

You can't deny recent and most powerful hardware is always made for PC first, so PC is the most powerful hardware when you want some reference.
All games are developed on PC , more games comes from console to PC like SoulCalibur , RE2, Devil May Cry 5 and others. They are developed on PC indeed.
The best versions is always the PC,4K 60fps, more details on scene, higher textures and better shadows than the console version.
Same for Far Cry games, they are developed on PC , and the best version is PC.

Anyway, good luck with your discussions on the forum LOL

 

Yes, it does matter, because it affects the type of game you can create. If the game you want to create will be unusually heavy on resources and you haven't planned on bringing it to console from the start you might not be able to bring it to those platforms.

There aren't too many examples of this though as most people aren't developing resource heavy games and those that do are almost always indies. Space Engineers is one of the few examples I can think of. If you try to load anything other than a small map on a system with less than 16 GB RAM it gets about halfway through the load, runs out of memory, and crashes.

Optimizations might help a little bit but when you're dealing with an enormous voxel-based game world you just need memory and 8 GB is simply too low for anything other than a small world. No Man's Sky only achieves it by constantly regenerating the world from scratch which has its own set of limitations.

https://store.steampowered.com/app/244850/Space_Engineers/

 

Things that have to brute force through computations like this are honestly ridiculously rare. Like, they number in maybe the dozens. We have not been hamstrung by computational power in any meaningful way in ages.

 

Tell that to the AAA games ported to switch ...

 

To be fair though we're still within the normal lifespan of a console generation. The PS3 lasted almost exactly seven years while the PS4 is only about five years old now and it had a minor refresh which will likely push the total lifespan to eight or more years. By the time we reach that we will be hamstrung by computational power.

 

And by then, consoles will be dramatically more powerful as Microsoft and Sony push for "true 4K rendering" and Nintendo finds a bold new way to bring back the Wii vitality sensor.

You can quote me on "True 4K." We're going to be hearing that a lot when the next generation is announced.

 

I think @Murgilod can't argue on that anymore

PC is already 4K 60 fps, you can even display a game in 8K on PC what PS5 or next Xbox won't do.
You are trolling, it's outstanding LOL

 

I'll laugh maniacally if it's "Upscale 8K".

 

Here's the thing, champaroonee:
Most games on PC, even recent releases, don't even support 4k. A lot of them don't even breach 1440p. A lot of them are fundamentally incapable of hitting 4K. 8K is essentially not even a reality at this point. And even if it was, it wouldn't matter because cross-platform releases are overwhelmingly optimised for consoles. Hell, most PC exclusive games aren't even exclusive for power reasons, but for input ones. The rare exceptions are things that are pushing huge voxel worlds, and even that's kinda funny because voxels terribly inefficient.

 

I would love to debate consoles as much as the next guy, but we should probably do it in a separate thread. The console vs PC debate is pretty off-topic for this thread.

 

Okay, let's all take a step back before we say something we'll all regret and ask ourselves one thing first:

Given the thread itself and its premise, does it really matter?

 

Hahaha you made my day LOL

 

Aside from over 7,000 games coming out on Steam in 2017 alone, this is literally two games. Even in the AAA space, you're lucky to see above HD resolution though.

Because barely anyone owns a 4K display.



And the cards that can drive a 4K display at any reasonable frame rate start off at around $350.

4K has limited support because it turns out that having to spend $700 between a monitor and a graphics card is not a market that has loads of support.

 

Plus, I think even those that can afford it, generally tend to prefer a 1080p high frame rate monitor as opposed to 4k.

 

Goal post are moving that's nice!

However the main complain of AAA dev, this gen, is how weak the jaguar processor had been, especially for open world. Ambitious game like anthem paid that price hadsomely (you can feel the devlopment pain and the cut contant in the final version, literally with out of order character dialogue). In fact, a lot modern console game are 30fps because they are cpu bound. Especially in the age of open world game with plenty of life interaction to have,where you need to have ginourmous amount of raycast on complex geometry. Which is the main discussion around the next gen, ie expecting a big cpu boost to increase gameplay. Also storage limit and bandwidth are starting to be big big problem.

 

How is it moving goalposts to point out your claims about 4K and 8K support make no sense at all? Also, we have seen more open world games this gen than literally any other, to the point where we're starting to see genre fatigue. On top of that, Anthem's failures were less performance based and more "so EA kinda got their fingers all up in this and also almost all of the best talent at Bioware is gone."

 

It's not my claim, I'm not the one who start arguing...

But 4k support is what the xbox one and the ps4 pro actually does, that it make sense or not is not my concern and is an opinion. Also there is a dedicated based on enthusiast building machine pushing 4k 60fps on game and technique like checkerboarding are used to cope with 4k exigeance. So there is a market for 4k whether you and I like it or not, and the size of that market don't matter at all as long as it served.

Genre fatigue is equally an opinion, and while it might be true, it still not relevant to the fact that those game where held back by the cpu power. In fact a lot of fatigue in a genre is due to how stagnant formula of a given genre become (thanks ubisoft) not thegenre itselff, but that's my opinion.

And finally public perception of what's the problem of anthem is, is not the same as the dev problem of what anthem had, which a lot can't be said because NDA and career building soft talking (like "i'm glad to leave that awesome team in which we made so many great game, because that make sense"), but the main things people with contact with insider at bioware is not ea (only) meddling the game, but ambition (making the bob dylan of video games) translated overtime internally as "making game is hard".

IN fact I think the final product is a case of a lead stepping up to save the trainwreck and apply given formula tacked on the game to save the product by making things done. I hadn't keep up with people I know in the industry to have more juicy bit tough, so I'll stop at this. I hope it will trickle over time like andromeda (which wasn't due to ea meddling only (frostbite engine mandatory) but aimless goal and vision (pcg world story based game, not for the faint of heart following buzzword) and lofty target (hundred of planets, that's30mn per planet give you 50h of game, that's kind of heavily dilluted rush through content for a story base game) which prompt a heavy course correction on a very short time ... just like anthem it seems but pull more competantly.

 

There is a market for 4K (especially in gaming spaces) much like there was a market for 3D goggles. It is not based on consumer demand, but rather guiding toward what is thought to be a lucrative business venture based on the hopes they can convince people to buy something they never actually had any inclination to. This has been a running problem in tech for over a decade.

Genre fatigue is a collection of opinions and a known factor not just in gameplay systems but fiction as well. It's a pretty damn common thing and is a huge part of why we stopped seeing disaster movies, for instance.

Anthem is underperforming because the game is poorly designed and a technical disaster. These things stem from its development and the issues surrounding it, including a lot of issues internal to both EA and Bioware. Where is this ambition you speak of? Because I sure as S*** didn't see any of it in the actual game.

 

Like there is a market for VR even if it is tiny compared to majority of Steam gamers that does not have VR hardware for example. Anyway stubborn people can't understand that LOL
Have a nice day.

 

That's what I call dev problem, they were cut and reformed

We have well know case of that, but it ended well, the first assassin's creed. They had a lot of ambition and none turn out as compelling gameplay, until someone got call to take the lead and make ad hoc traditional mission structure that made it into a very repetitive gameplay, in 3 month, before shipping. The sequel turn out to build a better game building and exploiting whatever innovation survive the big chop. The lead of assassin's creed what spectacularly fired after that.

A failed example is andromeda, they wanted a huge pgc open world building on mass efffect 1 mako mission, half the team couldn't make pgc worked, the story team couldn't build a well paced story with that template, so they retrofittted the game into a traditional template of the previous game in two years after 3 spent, which stressed develoment and produced a game that couldn't live up to the ancestor. None of the ambition survived except a better fighting gameplay loop and a decent mako driving (without the level design to take adventage of it).

Anthem ambition is hinted at the lore, the "anthem" left by the precursor can shape reality and they wanted a dynamic world where you had atat riding merchant going from multiple city, plenty wildlife interaction (like monster hunter) in a seamless world, huge battle, all things that basically plays on current generation weakness (cpu is weak to handle wildlife and lot of physics, lush vertical forest with reality bending events stress the streaming of scene). They probably had those thing working in isolation but not together (you can see hooks in the lore and also the gameplay and level design to explain these addition, but also in the type of hack they employed (mismatch between recorded mission statement and actual mission objective) to make it work with ad hoc new gameplay). Basically the gameplay took a hit and they had to resort to ad hoc last save throw by making basic mission structure to make it shippable FAST (ie reusing thsame copy paste mission structure and objectives).

And that's actually really common but not to this extent in all game dev, especially AAA one or dev tied by deadline (either time or money), you have to ship the game at some point. I know it happen to the witcher 3 where they literally expedite two third of the game to fill the world, though they had found good process to meet a good quality bar relative to the competitor, it happen to outcast the 1995 voxel game, where the main graphist told me there was no time to properly animate the main character that's why the walk animation is stiff, etc ... that's the natural course of game dev, but with the high stakes of blockbuster game failure is more itchy, anthem hadn't pull it off.

Basically the hardware couldn't met the demand of the ambition of the game so the ambition were dropped....

Which bring back nicely some theme in our discussion, some gameplay can't be made without adequate hardware (here is cpu and storage bottleneck), anthem was envisonned as a living an breathing seamless mutable world but even the ATAT plagiarized vehicule couldn't move anymore in the world. Turns out that making high speed flying character (which solve some movement problem open world problem) with the ecology of monster hunter can't be handled by current hardware. Monster hunter can pull it off by having slow characters on ground (ie lots of visual blocker) and highly segmented zone, and still have long loading time. They had to compromise, you always have to compromise somewhere. Anthem didn't compromise fast enough and end up with a correction course that led to a subpar shippable.

Now people who correctly planned for ambition are the warframe folks. Tipping my hat to them, but even then their own start was quite rocky.

The problem with market and success is that it's literally random, AR and geolocalisation game simply didn't work outside a niche, then pokemon go happen. Sometimes old tech don't work in the past, but suddenly work for whatever reason or events. I'm no fortune teller so I'm not commenting on that other than stating what's happening now.

Genre fatigue is more like an effect of follower flooding the market with half baked product. Platformer is tired unless you are nintendo, open world was already tired until BOTW and Horizon zero down kick the interest up for a while. Battle royale clone felt tired after fortnite very fast, then apex happen ... I don't blame the genre, I'm blaming people saturating the market with half hearted attempts at reproducing a formula.

 

Found an article on AMD planning on starting to stack memory onto the CPU -> https://forum.unity.com/threads/amd-to-3d-stack-dram-and-sram-on-processors.646705/#post-4333459

It also includes the potential of chiplet technology so CPU/GPU and other chiplets can be placed onto the processor die and then HBM or DRAM can be stacked onto the processors.

 

And? Are you just posting this here because your old thread got closed for having nothing to do with anything.

 

I posted it in a new thread because I thought the potential boost in performance we will get from this new development could be a game changer.

However it highlights the trends I mentioned here and the changes to the industry are already on the way.

 

No, it really doesn't. This is a natural evolution of current technologies. This has nothing to do with your weird future tech stuff that's been debunked.

 

AMD's subreddit community had been theorizing that this would be coming down the pipeline after they showed off the new chiplet approach being used by Zen 2. Plus we basically need it now as the next generation of memory is at least a year away and the current AMD APUs (Zen+) are memory bandwidth starved for most games with dual channel DDR4.

 

AMD has already done that with some GPU, so it makes sense AMD would also do that with CPU and APU as well. One thing to note is that this development will probably lead to more consolidation into the CPU offerings instead of new dedicated hardware cards. Basically the development would likely indicate the opposite of what you initially advocated in this thread.

There will be a point during the next few years where APUs get powerful enough to displace nearly all use cases for GPUs. There will be even less need for dedicated hardware cards.

 

Well, this has strayed well off its very thin topic in the first place.

This is a practical discussion site for people actively engaging in game development or using unity. Today. Now. Not for random speculation of non existent or future hardware. One would think that is a simple enough concept to understand. One is sadly disappointed.