CPU vs PPU vs GPU - Who has the future of physics?

[azopi]

The importance of Physics will grow in the near future without a doubt. PS3 has already done the job with Cell. Remember the day when the first PS released. The bunch of cross-platform (PC and PlayStation) titles made the 3D accelerator (which later evolved to a GPU) a must-have part of a Home PC.
Now the situation is similar. PS3 is coming. Developers who want to port between the PS3 and PC (and XBox) platform need some extra power in the PC to achieve the goal.
But who's got the power?
Who will do the Physics homework?

First candidate: the CPU.

Due to the weak parallelism CPU has and to the fact that it has other tasks to do (like AI, gameplay, control etc.) it's unlikely that today's CPUs will succeed on this area.
Multicore CPUs can help but those are still far from reality. According his long-term roadmap Intel plans Cell-like CPUs to be released only somewhere in the first part of the next decade. Check it here:
http://images.anandtech.com/reviews/tradeshows/IDF/2005/Spring/Day3/FutureCPU/evolution.jpg
http://images.anandtech.com/reviews/tradeshows/IDF/2005/Spring/Day3/FutureCPU/vision.jpg

I think it will be too late since the physics-stressed applications and games needs the extra power right now.

Second Candidate: the PPU

As we already know, Ageira has just introduced the first dedicated hardware for physics (the PPU) on the market. It has been shown to a small group of developers on GDC. It uses the Novodex SDK, so games using this engine can take advantage of the hardware acceleration right now.
The problem with PPUs is that it has to start from a zero user base. And who on earth would want to start a costly development to the small market of early adopters?
If the developers write the code to get the HW acceleration fully utilized than it won't run on a config lacking a PPU. If it doesn't, consumers won't experience so significant difference in quality what could sell the PPU.

Third Candidate: the GPU

GPUs are well paralellized for physics tasks. But it suffers from the lack of functionality and programmability . I know, it's historically getting more flexible with the advents of each new generation of hardware and shader model, but it's still far from the needs.
If the GPU can evolve to a general purpose streaming processor in the not too distant future, it can win. Simply because high- or mid-level GPUs are working in every gamer's / home users config. After Longhorn even office workers may have a powerful GPU.
So if GPU-makers get the API support (from Microsoft of course) and they succeed to do the job efficiently, the coming support from developers is unquestionable.

The race is getting started.
What's your bet?

[Alejux]

You forgot the 4th option: Motherboards.

[azopi]

You forgot the 4th option: Motherboards.

Motherboards with integrated PPU? It's almost the same as the second one :)

MoBo makers love to keep their products cheap. They won't integrate yet another expensive part onto the board increasing the cost of manufacturing. It could only be an option for the enthusiast market, but hardcore gamers prefer discrete parts.

Yes, we have a lot of integrated GPUs, but keep in mind that those are in the lowest segment of the market. Everyone need a GPU, but office workers won't demand real-time physics simulation to boost their word-processing, accounting and logistics performance ;)

[Alejux]

I'm not saying it's going to happen. Just that you forgot to mention it, specially since there has already been some talks about MoBo integration from people working on PPU's.

If you ask me, the reasonable approach, for the PC industry, would be to integrate it to the GPU. But it's just an uneducated guess, since I have NO idea on the manufacturing costs of a PPU element.

About CPU's...well I think Intel and AMD got off their lazy asses too late for this. It will take years before they can come up with something even close to a Cell and Xenon in the FP department. If you ask me, I think they wasted a lot of effort in making wordprocessors and office aplications as fast as they could without realizing that there's only one reason people buy really fast computers (besides DB servers), which is games and graphics processing.

[Martin Eddy]

I voted for PPU, but I'm hoping it will be added to next gen graphics boards as an option. Imagine how long the model number would be on a Gainward Product. :lol:

[Sandwich]

GPU? A chip that is so general purpose that it could be used for physics calculations as well as pixel pushing wouldn't be a "GPU" anymore.
It would be a vector processing unit of some kind.
Dedicated GPUs are here to stay. All consoles and gaming rigs need fast graphics and the most efficient way of doing that is through dedicated hardware.
Those pipelines shouldn't be idle and available for anything else while a game is being played.

I think the best bet for realistic physics in future games is a multicore CPU or a PPU.
I'm hoping for powerul PPUs, but I think it'll be CPUs. Games don't have powerful physics engines nowadays and it's not hard to imagine realistic physics will remaining secondary, used extensively only in a limited genre of games. I'm thinking Flightsims, racing games and shooters opposed to rpgs, platformers, strategy and puzzle games.

[MDolenc]

I'd go with GPU, becouse you are bound to repeat work with seperate PPU once more "exotic" surfaces start to appear. If there will be such thing as geometry shaders in DX Next, PPU will also have to support these.

[azopi]

GPU? A chip that is so general purpose that it could be used for physics calculations as well as pixel pushing wouldn't be a "GPU" anymore.
You are right. I mean under GPU the successor of today's GPU regardless what it is called. The predecessor of the GPU was 3D accelerator and VGA adapter. Did you imagine back in the 80's that there will be a time when that cheap connector between your motherboard and your monitor will be the fastest, the hottest and most expensive and important part of your PC? :D

It would be a vector processing unit of some kind.

Now I see why ATI calls its unit Visual Processing Unit. They don't have to change the acronym, just simply replace the first word ;)

Dedicated GPUs are here to stay. All consoles and gaming rigs need fast graphics and the most efficient way of doing that is through dedicated hardware.
Those pipelines shouldn't be idle and available for anything else while a game is being played.

I can't agree. I don't say that the pipelines would idle, but would it be impossible to add some physics tasks to future's graphics pipeline economically without a noticable performance penalty. AFAIK collision detection can already be done by the VS unit. And you could be able to scale down or up the stress of physics in line with the stress of graphics. The more detailed physical word can mean the more detailed visual world.
And if one GPU is not enough, there is SLI or Multi-Rendering or whatever. Plus there are rumours that GPU-vendors are going multi-core as well.
Plus you can choose the 16-pipe version instead of the cheaper 8-pipe model.

Graphics is traditionally paralel and scalable so if it's getting more programmable the post-GPU processor (you can call it general purpose stream processor) can be the best choice for physics and graphics as well (or even AI?).


I think the best bet for realistic physics in future games is a multicore CPU or a PPU.
I'm hoping for powerul PPUs, but I think it'll be CPUs. Games don't have powerful physics engines nowadays and it's not hard to imagine realistic physics will remaining secondary, used extensively only in a limited genre of games.
Believe it or not, with the advent of PS3 and Cell it will change very soon.

[nutball]

It'll end up in the GPU or the CPU... probably the CPU 'cos Intel/AMD seem to have run out of other useful things to spend their transistor budgets on, whereas ATI/NVIDIA haven't. It would also get the tech into the market in useful quantities (ie. in every new box sold) allowing games designers to target it. Hell, it might even be useful for other stuff (DSP, for example).

I honestly can't see an add-in solution working. I *can* see Ageia selling themselves to someone else.

[Ragemare]

I voted PPU, although 5 years is a long time, the PPU could have come and gone by then.

I think it's too early to say really, Ageia's success depends on how much of a difference it makes to UE3 games and the like. If it does make a big difference, then it could get a foothold in the enthusiast market and then maybe a more mainstream one.

[DOGMA1138]

i doubt we`ll see PPU on g.cards any time soon.
and when we do it prolly would come from the card maker side and not the gpu makers.

[Chalnoth]

CPU.

Why? Well, the GPU just isn't going to be available for physics calculations in games. Put simply, graphics takes a whole lot of processing power, and GPU's are only going to get better at keeping all of the core active at processing graphics.

I just don't think the PPU itself will get off the ground, as it'll be too low-volume to launch at decent prices, and thus will be expensive, and people just won't buy it without a killer app. And I personally feel that, even if it would be significantly better than what the CPU can offer, it'll just take too much work to design that killer app to get the market started.

Now, that leaves the CPU. CPU's have, for some time now, been shipping with units optimized for limited amounts of parallel processing (SIMD stuff), and will soon be shipping as multi-core systems. Eventually they'll probably all go the route of Cell, with a small number of powerful single-threaded CPU's, and a large number of CPU's designed specifically for parallel processing. In five years, we'll likely be up to the stage where we're just starting to see multicore move away from multiple identical cores, and as such the CPU's of the time will be well-suited to doing physics, and will be getting better at it all the time.

[jvd]

THe cpu .


why ? Multi core cpus . The quickest way to get more performance is to dump the overhead of dx onto a second core which i believe one developer (mabye sweeny ) said was up to 30% (not sure if the number is corect).

I think ppus will have a short life unless they can bring alot of power for sound card like power . Its hard to sell physics

[ZoinKs!]

I voted shared between cpu and gpu; probably mostly done on cpu. But I hope PPU's are awesome enough to make themselves worthwhile. Maybe they could do physics and sound, I don't know but I can hope.

[ShootMyMonkey]

I really don't have any faith that the GPU will ever be programmable *enough* to be viable for full-scale physics system processing. At best, it might help speed up the solver processes, but all the data shuffling will probably override any advantage that comes from the higher raw computational power. That's the general reality of any and all general-purpose tasks shifted to the GPU. Maybe one day, memory will be fast enough both in latency and bandwidth, and our connections between CPU and GPU will be fast enough in kind... but I don't expect to live to see that day ever come. Not that I don't think technology will advance quickly... I'm just that cynical.

As for the PPU... well, the PPU we know right now is essentially an API on a chip (namely Novodex). When you limit the developers to a particular API, that's not going to be suitable for everybody. The problem with an implementation like that is that from game to game, you can't keep the physics really identical. Even I, _not_ being a physics programmer, know that much. You have to be ready to make certain exclusions and inclusions from title to title because characters and worlds will have certain different demands and altered features. For that matter, physics systems evolve and people find new ways of doing things. Now if the PPU was little more than a solving engine to accelerate things like LCP reductions and raycasts and other similar things which are more or less universal, it would be a different story. That can be utilized as the developer sees fit from project to project and simply become a way to do your calculations really fast.

In the end, I have to vote CPU given the way things are. It's the only thing general enough to give you everything you need except for performance. But between getting twice the performance for half the generality, that's not a viable tradeoff... I'd rather take half the speed and work on speeding everything else up.

[Sxotty]

I vote PPU.

Of course this will only happen if it ends up in future consoles to give it enough game support and economy of scale.

[SlmDnk]

GPU is dedicated for graphics and PPU is dedicated for physics. Look where we are now with GPUs, without would be just horrible. I hope/voted for dedicated PPU.

Here's my AGEIA PhysX PPU info page once more:

http://personal.inet.fi/atk/kjh2348fs/ageia_physx.html

[Brimstone]

I voted PPU.

The PPU faces an uphill battle without a doubt, but if Ageia gets in into a console the PPU's future looks much brighter. They must be desperate to get into into the X-Box 360, so they can establish a marketing/developer beachead. If they fail to achieve that, the task of getting it to become a de facto standard like a GPU will be much more difficult. Cetainly full support from Epic with their Unreal Engine 3.0 is a big boon regardless what happens in the console arena. Courting ID's John Carmack would be a wise move. That's assuming Mr. Carmack likes the idea/concept.

It won't be an overnight revolution, but with smart marketing decisions, a PPU does have a fighting chance.

[Chalnoth]

GPU is dedicated for graphics and PPU is dedicated for physics. Look where we are now with GPUs, without would be just horrible.
Except the two aren't directly comparable. In graphics processing, there's a very large amount of work that needs to be done that is both low-precision and repeated many, many times. The primary example is bilinear filtering. This was really the first thing that separated 3D acclerated rendering from software rendering.

This being a major difference really is what made 3D graphics worthwhile to start: once you modified your game to support 3D graphics, you automatically got higher resolution and bilinear filtering essentially for free. Any support for 3D acceleration, then (as long as it was not horribly buggy), resulted in a huge increase in quality.

You just don't get that with physics processing. Not only is it much harder to sacrifice precision for most operations, but I'm just not convinced that a PPU can do things that a CPU cannot do. I'm even less convinced that it'll be able to do things that a dual-core CPU cannot do.

I will grant, of course, that dedicated hardware will always be faster. But my question is, will it be noticeable?

[The Postman]

Definately CPUs. They're already in place and are going dual-core soon.

[Alejux]

Dual core CPU's won't be able to handle physics nearly as well as PPU's or FP monsters like Cell. It will take at least 4 years IMO, for a solution from AMD/Intel with that kind of FP power. By then, I hope PPU's will already be popular and cheap.

But my opinion, is that they'll end up integrating a PPU into video cards. GPU's will eventually be known as Game Processing Units instead of only graphics one. Just a hunch.

[nutball]

It will take at least 4 years IMO, for a solution from AMD/Intel with that kind of FP power.

Why? This tech is not difficult (Google for GRAPE). Why 4 years?

[Brimstone]

I will grant, of course, that dedicated hardware will always be faster. But my question is, will it be noticeable?

I could imagine it being very noticeable in a game like Rainbow 6. With bullets impacting all sorts of surfaces/objects and explosions from grenades going off, the enviroment being blownapart and tossed around would enchance the sense of realistic violence. The gameworld will come across as being more alive.

[Goragoth]

I will grant, of course, that dedicated hardware will always be faster. But my question is, will it be noticeable?
I believe that the answer to that question is very much dependant on what the developers do with the extra power that the PPU will offer. Certainly the easiest thing would be to just have more "stuff" on screen when there is a PPU in the system (e.g. more debris flying about when something gets blown up). This would be pretty lame and won't sell physics to the masses.

If a big developer like Epic were to really get behind it and offer some features exclusively on systems with PPUs it might be a different story. I'm thinking perhaps having nice dynamic cloth animation. I have seen what a difference this can make in EQ2, it looks beautiful and there it is still very limited because the hit in processing power would be far to great otherwise. Now imagine that but with proper collision detection, more detail and different cloth types with different properties (in EQ2 all cloth seems to behave like silk). No give characters long, dynamically animated hair as well. The qualitative difference will be huge.

Games don't have powerful physics engines nowadays and it's not hard to imagine realistic physics will remaining secondary, used extensively only in a limited genre of games. I'm thinking Flightsims, racing games and shooters opposed to rpgs, platformers, strategy and puzzle games.
At first yes, but what games used 3D acceleration in the days of the Voodoo? That would have been flightsims, racing games and shooters. Just a thought.

Two things will need to happen for PPUs to take off.
1) Developer support. Not just having the ability to use the PPU if its there to make things a little faster/nicer but to really put some PPU exclusive features in their games that make people go "OMG WTF 1337!" :wink:
2) Open API support, either DirectPhysics from Microsoft or from someone else (OpenPL?). PPUs will not see widespread adoption by developers until this happens and there can't be any competition in the PPU market without it either. No competition is always bad and will mean that prices will likely stay high and innovation won't happen.

[Inane_Dork]

Well, if PPUs ever become an accepted part of the PC, obviously, physics will go there.

But since I never see that happening, I'm voting solidly in favor of the CPU. With multicore, the power will be there and the flexibility trounces the GPU's offerings.

[Alejux]

It will take at least 4 years IMO, for a solution from AMD/Intel with that kind of FP power.

Why? This tech is not difficult (Google for GRAPE). Why 4 years?

If you take a look at both Intel's and AMD's roadmap, you won't see anything greater then a dual core for the next years. There's nothing indicating that they will have a PPU capable CPU in the next 3 years (not counting the 1+ year time for it to become popular). If you ask me, I really think they wasted a lot of effort trying to make the fastest OOOE CPU's, which we all know is a dead-end approach, and not nearly enough effort in giving vectoring power like you find in CELL or XENON. The result is, next year, while AMD/Intel are selling expensive dual-core cpu's, capable of <50 gflops, Sony and IBM will be selling cheap CPU's that will be able to destroy them (performance wise) in 90% of the applications where CPU power is needed.

[Chalnoth]

I could imagine it being very noticeable in a game like Rainbow 6. With bullets impacting all sorts of surfaces/objects and explosions from grenades going off, the enviroment being blownapart and tossed around would enchance the sense of realistic violence. The gameworld will come across as being more alive.
Having a dedicated physics processor doesn't automatically give you an infinite amount of processing power to handle. And you also have to deal with the graphics side of things. So it won't necessarily give enough of an improvement to actually give games a noticeable boost.

[Sxotty]

You know what they could do:?:

Well here is an idea that spawned in my head thanks to the "courting John Carmack" bit.

Why not give Novadex to developers for free to integrate into their games, *if they program support for the PPU* Then it would really help their hopes I think.

Edit: Looks like they already realized this and are doing it themselves :P
Ageia bought the Novodex physics API and has since been giving it away to developers. Several upcoming titles use the API, but from what we've seen, they don't include it because it's better than Havok or any other physics middleware. They do so because it's optimized for multiple threads, which multi-core CPUs of upcoming PCs and game consoles will take advantage of. But that seems to be its one big trick. Pretty much everything else it does, all the other software physics engines do, too. If your game uses Novodex as its physics API, that doesn't automatically guarantee acceleration by the PhysX chip.

[Brimstone]

I could imagine it being very noticeable in a game like Rainbow 6. With bullets impacting all sorts of surfaces/objects and explosions from grenades going off, the enviroment being blownapart and tossed around would enchance the sense of realistic violence. The gameworld will come across as being more alive.
Having a dedicated physics processor doesn't automatically give you an infinite amount of processing power to handle. And you also have to deal with the graphics side of things. So it won't necessarily give enough of an improvement to actually give games a noticeable boost.

Well 125 million dedicated transistors fabbed on a .13 nm process burning 25 watts is going produce a lot of physics processing power. Current x86 CPU's are already struggling with heat issues, I just don't see how any Intel/AMD cpu coming out in the near term is going to come close in performance.

Even when Intel/AMD start producing a dual core.

a. The cores won't be anything like the PS3 CELL processor. The x86 enviroment doesn't demand a radical change with massive floating point performance at this time.

b. Dual core will be very expensive at first, and even then those dual cores aren't neccessarily going into hardcore gamers PC's. A lot of them will going into servers and workstations. It will take a while for dual core to become common in gamer PC's.

On the other hand the PhysX board will probably hover around the cost of a decent GPU. I doubt many gamers are going to balk at the price. Also since its a card, when they do upgrade to a new CPU, just like a GPU they can remove the accelerator and install in the new system.

Having games that really tax the PhysX chip is important obviously. If the thing is overhyped too much before developers have had a chance to figure out ways to take advantage of the power, gamers could become disenchanted. Next gen console games on the X-Box 360 will push physics harder than any games before. For example, I can see John Madden Football and all of EA's sports games having a robust increase in physics processing. Most of those games will show up on the PC.

[Chalnoth]

Well 125 million dedicated transistors fabbed on a .13 nm process burning 25 watts is going produce a lot of physics processing power. Current x86 CPU's are already struggling with heat issues, I just don't see how any Intel/AMD cpu coming out in the near term is going to come close in performance.
Not necessarily. Firstly, CPU's run at much higher frequencies, which can help to compensate for the lack of transistors dedicated to processing. Today's CPU's also support SIMD operations that further improve their performance.

That all said, I'd be willing to bet that the PPU in question still is more powerful, but I just doubt it's enough more powerful to really be obvious to the gamer.

Even when Intel/AMD start producing a dual core.

a. The cores won't be anything like the PS3 CELL processor. The x86 enviroment doesn't demand a radical change with massive floating point performance at this time.
Not when they start, obviously, but later it makes lots of sense. It's not even about massive floating-point power, necessarily, but about the simple fact that it's vastly easier to increase parallelism than it is to increase IPC while keeping things single-threaded. And thus it becomes very easy to add in a ton of simple, low IPC cores designed for parallel processing. But there needs to be a shift to multiprocessing for most applications before this becomes useful, so it won't happen for a little while.

b. Dual core will be very expensive at first, and even then those dual cores aren't neccessarily going into hardcore gamers PC's. A lot of them will going into servers and workstations. It will take a while for dual core to become common in gamer PC's.
Possibly. I think you'll get a lot of enthusiasts jumping on the bandwagon just like people seem to be doing with SLI. I mean, SLI doesn't always increase performance by a lot, and it isn't really cost-effective to buy a new system sporting SLI, but it does seem like quite a few are going for it, or, at least, for motherboard supporting the tech.

On the other hand the PhysX board will probably hover around the cost of a decent GPU. I doubt many gamers are going to balk at the price. Also since its a card, when they do upgrade to a new CPU, just like a GPU they can remove the accelerator and install in the new system.
$300 for an improvement that might be had with simply a faster CPU sounds like a lot to swallow to me.

[Brimstone]

The "claims" by Aegia predict up to a 100 fold increase in performance. My guess is that a more interactive game enviroment will be noticeable. Right now in many games the enviroment is very static. Gamers seemed to have reacted positivly to Half Life 2's physics.

Computer Games - Since the announcement, some people (including some game developers) have expressed skepticism that the average PC gamer would spend extra to have a physics processor on their PC. What is your response to this?

Curtis Davis - If you purchase a new GPU, you will experience 2X or 3X performance improvements, whereas with a PPU you can expect to see 100x performance improvement in effects and more interactive game play. Additionally, it’s about the titles. We will have at least five titles that matter that will take advantage of the PPU and hard core gamers want the best experience possible. Also, this will be available to motherboard and laptop manufacturers as well as add in card vendors.



Also from the same interview.

Computer Games - You are using the NovadeX physics engine as the SDK for the physics chip. Is it possible that other software physics engines could use your technology in some way?

Curtis Davis - Yes. We are engaging with other software physics engines. Currently, however, games can be designed with a combination of NovodeX and other physics engines.



http://www.cgonline.com/content/view/451/2/


So I guess Duke Nukem Forever physics could be accelerated even though its built upon something different to NovodeX. Although it wouldn't tax the card that hard in all likelyhood.



Madden 2005 has some new physics that people have taken notice of. So by just scratching the surface the press is raising their eyebrows, imagine with dedicated hardware and complex interaction..

Gamespot is showcasing some incredible footage of Madden NFL 2005. The new collision detection gives the players realistic weight. Hits look really painful. And Madden’s voice has never sounded so irritating. Check out the game clip and keep your eyes open for how the players shift their center of gravity as they run. You can say what you want about the Madden series, but they certainly know how to make your jaw drop year after year.



http://www.joystiq.com/entry/7154052284095717/

[jvd]

I just can't see this doing that much that it will sell at 300$ , 200$ mabye , 100$ yea i'd say it would catch on like wild fire even if it only increases your performance by 10fps in most games .

No i think dual core cpus will drive physics . The price of dual cores will continue to go down and they will continue to get faster and the installed base will scale much quicker than this physics card as in a year or so down the line 90% of shipped pcs will have a dual core processer .

[Brimstone]

Not necessarily. Firstly, CPU's run at much higher frequencies, which can help to compensate for the lack of transistors dedicated to processing. Today's CPU's also support SIMD operations that further improve their performance.

That all said, I'd be willing to bet that the PPU in question still is more powerful, but I just doubt it's enough more powerful to really be obvious to the gamer.



Comparing the clock speed of a PPU to a CPU, will be just as effective as comparing a GPU's clock speed to a CPU if the algorithms are highly parallizable. The architechture still hasn't been disclosed. Who knows what breakthroughs Aegia may have on the chip?

[Ragemare]

If a big developer like Epic were to really get behind it and offer some features exclusively on systems with PPUs it might be a different story. I'm thinking perhaps having nice dynamic cloth animation. I have seen what a difference this can make in EQ2, it looks beautiful and there it is still very limited because the hit in processing power would be far to great otherwise. Now imagine that but with proper collision detection, more detail and different cloth types with different properties (in EQ2 all cloth seems to behave like silk). No give characters long, dynamically animated hair as well. The qualitative difference will be huge.


These are the lines I'm thinking along. Gabe.N said that the Source engine can easily handle cloth/hair simulation, but it simply took up too much CPU time. So the difference between haveing one of these things and not haveing it, could be the difference between haveing cloth simulation and not haveing it, or haveing it, but with worse performance.

It could be the difference between haveing high definition fluid dynamics and low definition fluid dynamics. If developers look I'll think they will find lots of places where they can create work for this thing without createing too much more work for the GPU.

[Alejux]

Not necessarily. Firstly, CPU's run at much higher frequencies, which can help to compensate for the lack of transistors dedicated to processing. Today's CPU's also support SIMD operations that further improve their performance.

That all said, I'd be willing to bet that the PPU in question still is more powerful, but I just doubt it's enough more powerful to really be obvious to the gamer.



Comparing the clock speed of a PPU to a CPU, will be just as effective as comparing a GPU's clock speed to a CPU if the algorithms are highly parallizable. The architechture still hasn't been disclosed. Who knows what breakthroughs Aegia may have on the chip?

I agree with you. Saying that a dual-core CPU will do PPU work, like the Novodex, is like saying you won't need a GPU anymore either. We all know that it doesn't work that way. From what I read, Algea's PPU is a highly parallizable piece of hardware. It's possible that a CELL chip might achieve the same kind of results, but I seriously doubt we will see anything close to that level of performance in a dual-core Opteron or Pentium.

[DegustatoR]

If you take a look at both Intel's and AMD's roadmap, you won't see anything greater then a dual core for the next years.
Quadcores will be out by the end of next year.

http://images.anandtech.com/reviews/tradeshows/IDF/2005/Spring/Day3/FutureCPU/evolution.jpg

http://images.anandtech.com/reviews/tradeshows/IDF/2005/Spring/Day3/FutureCPU/vision.jpg

Edit: And yes, i voted for CPU+GPU option. Just b/c there will always be some physics-related things which can be done on a GPU quicker and nicer than on CPU, although CPU will remain the only GP-processor in the system for many years to come.

[Alejux]

If you take a look at both Intel's and AMD's roadmap, you won't see anything greater then a dual core for the next years.
Quadcores will be out by the end of next year.


From what I heard, AMD/Intel will be ready to fabricate quad-cores in 2007. Put 1-2 years before they are on the market , popular and cheap, and you're talking 3-4 years, like I predicted. Having US$ 800,00 quad-cores on sale , will not influence the PPU market.

[Inane_Dork]

From what I heard, AMD/Intel will be ready to fabricate quad-cores in 2007. Put 1-2 years before they are on the market , popular and cheap, and you're talking 3-4 years, like I predicted. Having US$ 800,00 quad-cores on sale , will not influence the PPU market.Quad-core is coming a heck of a lot faster than that. Too bad all my info is passed on by word of mouth or I'd link to it. :P

[Brimstone]

The PhysX chips is a first generation 125 million transistor core based on the .13 nm node. A second generation core targeting .09 nm at TSMC should double that to at least 250 million while having a modest gain in clock speed. It will be a repeat of what nVidia and ATI have done with GPU's over the last 5 years as far as performance goes. It will be Moores law cubed all over again. Ageia will have lithography on their side and inovations in architechture as a result of sustained R & D. I just don't see multicore x86 CPU's that will be arriving over the next few years coming close to outperforming dedicated hardware when it comes to in game Physics.

[Alejux]

Has anyone heard any real news on pricing?

I don't see PPU cards becoming successful unless they're in the $100 - $200 range.

[DudeMiester]

PPU, but probably integrated into the soundcard.

[Karma Police]

Well, having the UE3 engine supporting the dedicated PPU certainly makes a difference, I would think.

All those who say the PPU will be still-born probably think the UE3 engine won't be that important, i guess.

:?

[Inane_Dork]

All those who say the PPU will be still-born probably think the UE3 engine won't be that important, i guess.

:?Well, that's not true in my case, anyway.

A popular engine supporting a feature does not necessarily make the feature commonplace.

[Chalnoth]

All those who say the PPU will be still-born probably think the UE3 engine won't be that important, i guess.
Just having support in the engine is not enough. The game using the engine must also have an option to enable extra physics for use with the PPU. And this option must make a very noticeable difference to actual gameplay, without adversely affecting performance by large amounts (f.e. from extra graphical load). This is what I doubt will happen.

[Karma Police]

Just having support in the engine is not enough. The game using the engine must also have an option to enable extra physics for use with the PPU. And this option must make a very noticeable difference to actual gameplay, without adversely affecting performance by large amounts (f.e. from extra graphical load). This is what I doubt will happen.

I certainly don't think they would make an engine that doesn't run well.

Of course anything can happen, but UE3 being made badly, I doubt it. :lol:

[Chalnoth]

Once again, it's not the engine I'm questioning. It's the content.

[Mate Kovacs]

AFAIK collision detection can already be done by the VS unit.
Are you sure about that? AFAIK arbitrary collision detection demands random access to the geometry data, which isn't the typical behavior for a VS unit. (Yes, there are vertex textures, but I think keeping the whole geometry data in vertex textures isn't a very common solution.)

[Remi]

While we're speaking about geometry...

Question: With "hardware accelerated physics" (whatever that means), will we let physics remained done against bounding boxes rather than detailed geometry? If no, how is another circuit (CPU or PPU) going to access that geometry which will be expanded inside the GPU?

(I tend to think that geometry expansion and physics should go together, on the same die. But then does that means splitting the GPU or integrating it all in it?...)

[Chalnoth]

Are you sure about that? AFAIK arbitrary collision detection demands random access to the geometry data, which isn't the typical behavior for a VS unit. (Yes, there are vertex textures, but I think keeping the whole geometry data in vertex textures isn't a very common solution.)
Right, so what you'd probably do is just some very coarse pre-processing on the CPU, if you wanted to make use of the GPU.

[scificube]

My vote is with the PPU.

The difference is unquestionable if the numbers are to be trusted.

A more real proof of this is within the Beta Novodex demo on Ageia's cite. Buildings Explode and Big Bang will completely smoke your CPU to the tune of dropping your FPS to the 1-3 range. (x86) Add 3 more cores all explictly doing physics (not plausible in a game enviorment) you could expect 4-12 FPS if things scale linearly. If not satisfactory knock it up to 30 FPS with extremely simple geometry, no AI, etc.

The only way comparable performance could be achieved is to abandon the x86 architecture or place something of a PPU unit on an x86 core. Cell is a radical change, but the likely hood that AMD and especially Intel will follow suit any time soon is pretty slim. The likely hood that 125million + extra trannies for a local cache for the whole of the chip or for the PPU like unit alone is equally unlikely to me. It would have happened when GPUs came about. It didn't. At least if it did happen they could add such units per core to keep up with the larger sizes of PPUs as GPUs have grown in size over time. This is highly improbable to me.

CPUs are general purpose processors and thus CPUs have not specialized in all areas in the past nor do I think they will in the future.

Even Cell is not specialized on the level of just physics or math. It's a heck of a whole lot of floating pt power that can be used when applicable. This is the Cell's greatest asset and caveat to it simultaneously.

---------------------------------------------

Support for physics in games is unquestionably what will kill or propel the PPU.

I differ in thinking that content cannot be put into games that gamers won't clamour over and at the same time leave the PPU less in the cold until such a time PPUs are both integrated and stand alone parts in common place.

Aesthetics is a prime example.

Fluid dynamics used with water with a PPU or without if a CPU etc is not up to the task.

Interactive cloth same deal.

Hair simulations.

Interactable objects that are not made significant to gameplay at all...swaying leaves, bending branches, cans rolling around, particle explosions etc.

PPUs can take physics off CPUs entirely which would obviously reap benefits in other areas if nothing else but overall performance.

When and if PPUs should become common place and physics can seep into gameplay itself there are too many instances to name where this would be kewl.

I am remiss to think Tim Sweeney would integrate what is worthless..not use it...or use it badly. I'm just not sold on that one.

If physics interaction is already showing signs of copy cat syndrome, HL2 -> Doom3 expansion, I think it's fair for me to conclude physics is becoming a valubable part of the content of games.

[Bludd]

CPU for much of the same reasons as Chalnoth's.

[Karma Police]

My vote is with the PPU.

The difference is unquestionable if the numbers are to be trusted.

A more real proof of this is within the Beta Novodex demo on Ageia's cite. Buildings Explode and Big Bang will completely smoke your CPU to the tune of dropping your FPS to the 1-3 range. (x86) Add 3 more cores all explictly doing physics (not plausible in a game enviorment) you could expect 4-12 FPS if thing scale linearly. If not satisfactory knock it up to 30 FPS with extremely simple geometry, no AI, etc.

The only way comparable performance could be achieved is to aband the x86 architecture or place something of a PPU unit on an x86 core. Cell is a radical change, but the likely hood that AMD and especially Intel will follow suit any time soon is pretty slim. The likely hood that 125million + extra trannies for a local cache for the whole of the chip or for the PPU like unit is equally unlikely to me. It would happened when GPUs came about. It didn't. At least if it did happen they could add such units per core to keep up with the larger sizes of PPUs as GPUs have grown in size over time. This is highly improbable to me.

CPUs are general purpose processors and thus CPUs have not specialized in all areas in the past nor do I think they will in the future.

Even Cell is not specialized on the level of just physics or math. It's a heck of a whole lot of floating pt power that can be used when applicable. This is the Cell's greatest asset and caveat to it simultaneously.

---------------------------------------------

Support for physics in game is unquestionably what will kill or propel the PPU.

I differ in thinking that content cannot be put into games that gamers won't clamour over and at the same time leave the PPU less in the cold until such a time PPUs are both integrated and stand alone parts in common place.

Aesthetics is a prime example.

Fluid dynamics used with water with a PPU or without if a CPU etc is not up to the task.

Interactive cloth same deal.

Hair simulations.

Interactable objects that are not made significant to gameplay at all...swaying leaves, bending branches, cans rolling around, particle explosions etc.

PPUs can take physics off CPUs entirely which would obviously reap benefits in other areas if nothing else the raw performanc.

When and if PPUs should become common place and physics can seep into gameplay itself there are too many instances to name where this would be kewl.

I am remiss to think Tim Sweeney would integrate what is worthless..not use it...or use it badly. I'm just no sold on that one.

If physics interaction is already showing signs of copy cat syndrome HL2 -> Doom3 expansion I think it's fair for me to conclude physics is becoming a valubable part of the content of games.

ditto :)

[Tim Murray]

I still don't understand the giant PPU circle-jerk. Reasonable people, are the ones that say that a PPU will be the bestest thing ever also the ones who say Cell will supplant x86?

[DegustatoR]

I still don't understand the giant PPU circle-jerk. Reasonable people, are the ones that say that a PPU will be the bestest thing ever also the ones who say Cell will supplant x86?
Mac users? 8)

[IgnorancePersonified]

I think people are kidding themselves about the price/performance of the latest dual xeon and opteron systems and the ability of even dual core systems using the same old architecture's to compete with dedicated hardware. A $200(max US) addin card would compete with these (http://www.gamepc.com/labs/print_content.asp?id=x36o252)
and there's still debate?? that's $1000 per cpu ladies. If people haven't noticed - CPU performance has hit a wall lately. I am not putting too much faith in intels knee jerk dual core price annoucements nor AMD's. They seem to be in a spin battle royale. Too suport Intel Dual core you need a platform with a chipset that hasn't been released yet. AMD requires at least a new revision 939 pin mobo. That's a big upgrade in any ones terms.
a descreete addin card is needed in the short term only. Get Developer support, Get market penetration and hype, sell the IP to be intergrated in - name_your_addin_card be it VPU, CPU, PCIe Sound Card or stand alone PPU card.
This company doesn't appear concerned with wooing dribbling teenagers with fistloads of mummys cash right at this moment. They appear to be more concerned with getting developer support and thier own physics engine used in as many up and coming engines as they can + support in other engines. Sounds sound to me.

Oh yeah - give me a $200 PPU unit that I can move between compatible systems over intel/AMD's vapourware any day. +1 for PPU.

[Chalnoth]

AMD requires at least a new revision 939 pin mobo. That's a big upgrade in any ones terms.
I currently have a socket 939 motherboard. Wouldn't be a huge upgrade for me at all. Plus I'd get the other benefits that come with upgraded CPU power, other than games.

a descreete addin card is needed in the short term only. Get Developer support, Get market penetration and hype, sell the IP to be intergrated in - name_your_addin_card be it VPU, CPU, PCIe Sound Card or stand alone PPU card.
When has that ever happened?

[IgnorancePersonified]

"I" is singular and in this scenario is by no means a majority or even close to a high percentage of "Gamers"
When hasn't it happened?

[Chalnoth]

Anybody who was smart and upgraded to a new AMD motherboard this past year was sure to have purchased a Socket 939 system. Every enthusiast out there should be aware that this is the socket that is most future-proof, and thus most worth the money.

[ndoogoo]

I vote PPU.

What is the point going back to the ZX spectrum days were the CPU did everything. One of the reasons the Amiga was better than the Atari ST(besides OS) was because of all the dedicated hardware acceleration.

Freeing up more CPU resources is a good thing as it means developers can improve or add new things.

Just think of new Spilter Cell
Voice recognition for (selecting weapons etc)
Improved AI
Vastly improved physics (all objects can take damage etc)
Goodbye rigid clothes, let the cloth interact with the model
More real looking hair instead of Mr and Mrs hard gel


It probably cant be done yet, but i would like to sound, graphic and computation accelerators intergrated into the one core. PPU , graphic and sound cards get replaced by the media card.

Vertex\texture data will hold all the information for sound, graphics and physical properties. So if you build a brick wall it behaves like one. Making it more imersive is what its all about.

And to finish off. A PPU will be a new toy for the hobbyist coder to play with. :P

[Anarchist4000]

I'd have to go more the PPU route. Keep in mind vector processing could be used for a lot more than just graphics related operations. Even apps like photoshop etc that deal with colors could take advantage of it. So unless a GUI started dumping a lot of data back to the system it likely wouldn't work to well. Cacheing works well for the vector based math since it's almost always related to processing huge chunks of data so you could fetch the needed data in advance.

A PPU of sorts with 32 little vector processing units could do some serious damage when it came to crunching numbers. It's be similar to a video card but you wouldn't need all the RAM necessarily, none of the pixel processing capabilities, or IO related stuff on the chip. Even a 400MHz clock with that many pipelines would go a long way on math ops. It would essentially remove the need for any SSE/MMX related instructions on the CPU if those operations could be dumped onto the PPU. Keep in mind you typically know when you're about to get a chunk of data that would be using SSE/MMX or processing a bunch of data in parallel.

It would essentially leave logic operations to the CPU and heavy math operations/data processing to the PPU. Make a PPU like a second CPU inside a system and give it a hyper transport link of sorts and it could work quite effectively. On top of that it could be sold like an add on so only people that really needed it in their system could select to go out and get one.

[scificube]

Hey that sounds pretty interesting :)

I wonder if there isn't some higher level of specialization though?

I mean like physics calcs hardwired in so to speak.

If the whole of the PPU is like this wouldn't it then be difficult to reap the benefits you describe? (although I do think this is unlikely)

I bet some things are hardwired but there are enough vector processors to still do what you describe or at least to some point.

I really wonder though if math could be taken off the CPU to this level by a PPU --- in this case the CPU would essentially become an AI chip (or logic chip in general) ...with the benefit of bossing around the other system parts.

If this were to a possibility or the future reality it certainly reminds me of the Cell chips current structure...with higher specialization and separate parts unlike on a single chip. In fact I think this would be more efficient and could allow the x86 boys and PC setup to terminally infect the Cell's progress into the desktop market. I say that as I look at the Cell as something of a rapid movement towards "systems on chips" where as PC still enjoys spreading the load throughout the systems parts.

Just some random thoughts. If I totally fell off the apple cart well...the devil made me do it :twisted:

[Chalnoth]

A PPU of sorts with 32 little vector processing units could do some serious damage when it came to crunching numbers. It's be similar to a video card but you wouldn't need all the RAM necessarily, none of the pixel processing capabilities, or IO related stuff on the chip. Even a 400MHz clock with that many pipelines would go a long way on math ops. It would essentially remove the need for any SSE/MMX related instructions on the CPU if those operations could be dumped onto the PPU. Keep in mind you typically know when you're about to get a chunk of data that would be using SSE/MMX or processing a bunch of data in parallel.
The problem is that a lot of these types of operations are heavily bandwidth-limited, and thus would not be optimal for anything that doesn't have its own local memory. And even then, if your input and output don't have enough instructions between them, then PCI Express bandwidth is going to get in the way, and you'd just be better-off doing it on the CPU (or on the GPU, if your result is just going to be output to the screen).

[scificube]

A PPU would have it's own local memory would it not? ...of course not a fast as a cache...and sending the results over the PCI-e bus does seem to pose a serious latency problem.

Oh well...

[Jodi]

My 2c:

What about game servers? The PPU would make life so much more simple for the collision detection / physics managment of a multiplayer server. It might mean that you could actually have full physics 100 player servers (if it could handle the BW), and still have CPU to play with for other stuff, like more advanced AI, or audio stuff or whatever.

Client side, the PPU would have cloth/hair/water simulation, and maybe sound field calculation as its strengths.

I'm quite sure the fashion industry would like the abilty to simulate clothing on various bodytypes without having to bother with cutting and materials and models.

Dammit, its getting me all excited. People have to start thinking OUTSIDE the box. If the PPU gets the right marketing people, and killer apps, it will have the abilty to work.

[Chalnoth]

What about game servers? The PPU would make life so much more simple for the collision detection / physics managment of a multiplayer server. It might mean that you could actually have full physics 100 player servers (if it could handle the BW), and still have CPU to play with for other stuff, like more advanced AI, or audio stuff or whatever.
Yeah, might make more sense in such a scenario, but more for companies doing subscription-based online games. I doubt people hosting stuff like UT servers would have the money for a niche-market product.

[Anarchist4000]

When I said use a PPU to offload the work I meant more along the lines of batching tons of basic math operations. When you run into a bunch of vector data it typically all hits at once. You could literally let the PPU go through and crunch numbers while the CPU spends it's time on other work. When when you get done the CPU can take all the data and compare it or something. It could possibly even be set up with it's own access to system memory.

In game design when you start the physics calculations you do them all at once. With a PPU you could do all the math on the PPU, then the CPU could go through and figure out what to do with all the new data.

Yes there would likely be high bandwidth requirements but latency isn't neary as much of a problem. The CPU could still do some of the physics calculations but when you got into the really heavy physics situations where you could be batching thousands of objects at once it could be beneficial.

And for a niche market almost all games still have to do their own collision detection and physics calculations on the client so it could be an addon card to get more performance out of your machine. Even compressing video or audio files could see a benefit from it.

[OVERLORD]

With the advent of dual core x86 procesors from AMD/INTEL now upon us. How difficult would it be to combine CPU & PPU cores into a dual core CPU/PPU processor package?

We have already seen how for example NVIDIA & AMD pushed performance to a degree by applying similar manufacturing techniques in advancing CPU and GPU performance. NVIDIA having built harware T&L into original Geforce and AMD having moved memory controller onto K8 processor.

AMD previously having been dependant on their chipset partners and motherboard manufacturers have gained most in combining memory controller within CPU package and have already shown hindsight in developing K8 with dual core in mind from the outset.

Wouldn't dual core CPU's benefit more by combining CPU/PPU functionality than developing identical dual core CPU's as proposed currently?

Like Hyperthreading before it dual core advancements may not be fully realised and may lay largely redundant unless optimised by OS and applications to utilise it.

This might be a more specialised approach but current dual core cpu designs also appear to have pros and cons. Whilst also running @ lower speeds than current single core CPU's.

Current dual core designs only solve a few problems in single core CPU designs like heat and increasing IPC performance stagnation in both AMD and INTEL designs.

This approach appears rushed by both respectively due to the limitations currently in their manufacturing process and the wall they both have reached in increasing performance.

This reaction is similar to when Intel rushed P4 in retalliation to Athlon. P3 IPC was still ahead of P4 @ same clock speed and took considerable amount of time to show performance gains to justify acceptance. Even now Pentium M processors, a combination of both P3 & P4 technolgies appears to have been better realised.

What do you think?

[Chalnoth]

With the advent of dual core x86 procesors from AMD/INTEL now upon us. How difficult would it be to combine CPU & PPU cores into a dual core CPU/PPU processor package?
Difficult? No, but it just doesn't make much sense. It'd make much more sense for CPU developers to do what they've been doing all along: produce general-purpose processors. One major reason why is that the clock speed of dedicated hardware is typically much lower than CPU's, while at the same time requiring more transistors. This just won't work.