One GPU can NOT handle physx + video: benchmarks

Status
Not open for further replies.
I doubt that this "professsional software developer" knows much about market research, profitability, cost management and etc. because eventhough the performance benefit with the right GPU could be huge, that alone won't be enough to push developers to adopt PhysX soon.

Those are more political and management issues. That is not my area of interest.
I'm merely trying to explain what nVidia is enabling here, just as I was trying to explain what Ageia was enabling before. So it's all about the technology.
We all know that just having the technology doesn't necessarily guarantee a successful product. But I never saw much point in trying to discuss that sort of thing. The market is too unpredictable. However, since nVidia has much better marketing than Ageia did, and nVidia also has a very large marketshare, I think we can all agree that nVidia has a much better chance of making this technology successful than Ageia did.
 
GPU_Physics.jpg

http://benchmarkreviews.com/index.p...ask=view&id=187&Itemid=1&limit=1&limitstart=4

The benchmark speaks for itself.
 
Well, actually, the best solution is probably two or more GPUs.
A PPU is only so fast, and if your GPUs are fast enough, you'll get PPU-limited.
Looking at the 3DMark Vantage results, a 9800GTX is about 4-5 times faster than a PPU when processing (nearly) only physics[](and a PPU is about twice as fast as a quadcore).
So there is a point somewhere, similar...

You stop the FUD-mark in a pipe a smoke it, in case I wasn't clear, I am speaking about GAMEPLAY, not FUD-mark..the picture changes when doing real-world gameplay...

to the CPU, where the PPU fully bottlenecks the framerate.
Show me one instance off this, and I will consider your argument...untill then, it in the same box a FUD-mark....the never-never-land...as test whow it the GPU becommming bottleneck when there is to much particle physics going on.

I think this point is within reach with a fast multi-GPU setup, where the PPU would bottleneck the framerate more than GPU physics would (as GPU physics power increases aswell with multi-core).
Oh, really...based on what?
FUD-mark?

In REAL gaming (like UT3) the PPU is still king...not to bad for a 2006 130nm tech PPU wit a TDP of 28 Watt...GPU's look so like bruteforce spaceheaters compared....that is not impressive...that is sad.
 
But in real world gaming terms what you're ACTUALLY ASKING of a developer is to drop 20-30% of their render time which they then have to balance off against the rest of the game, so to keep to their targets suddenly they have to drop the shadow resolution, or drop the amount of post processing they're doing to keep frame rate up, this is just sacraficing one visual element for another. Even if they don't do that and we all suffer bad frame rates in the game (as proven by the benchmarks) then we all just have to drop the resolution to make up for it.

You are working from the assumption that framerate is the primary goal of a game.
This is not the case. For years, developers have been adding extra features that impact framerate, but improve the overal appearance and/or gameplay.
Physics is just another of those features.
They're just taking advantage of the added power of new hardware.

Where as the alternative right now for developers is this hardware (the CPU) sitting there getting less and less used as the number of cores grow but their game remains on 1 core, so the driving factor here is to use untapped power and have an overall net gain.

You're assuming that games are still singlecore. Most current-gen game engines are already optimized for multicore, so this assumption is wrong. Aside from that, at least Havok and PhysX, the two main physics APIs, are optimized for multicore. So even if the rest of your game is not multicore, the physics part will be, if you use either library. So your argument is false. The only untapped potential is because games are 'neutered' to run well on low-end CPUs.
If you run the CPU benchmarks included with Crysis on a quadcore at very high physics, you'll see that the CPU is absolutely not spared at all. You're still CPU-limited once that rocket launcher starts blasting.
Same with 3DMark Vantage, it dumps a shitload of physics on all 4 cores... It even adjusts the number of objects in the test to the number of cores you have.
So the problem isn't at all that we can't use quadcores or anything.

Doing physics in the GPU is just shifting the workload around, thats all, it doesn't net is any overal gain.

It does, because the cost in framerate is smaller on the GPU than it is on the CPU. Aside from that, the CPU is freed up for other tasks, that the GPU wouldn't be able to do (efficiently). So developers can add more features to the rest of the game.
It's all about getting the most from the hardware. Arguing against physics acceleration is about as useless as arguing against graphics acceleration.
 
EPIC!

I emailed Gabe the other day about this and got a response, presumably from gabe...

Here's my email

Gabe,

Just a quick one, I know you’re a busy chap.

Any plans to use GPU based physics in the future? You’ve discussed your dislike for PPU based physics in interviews but since the GPU is a requirement for gaming to begin with, might this be something you use in the future? And if you have time also briefly elaborate on why?

Recent benchmarks show a fairly hefty hit to frame rate when running additional physics on the GPU, you might find this an interesting read if you’ve not already seen it http://techgage.com/article/nvidias_physx_performance_and_status_report/1

Thanks for your time,

Sam

And the reply I got.

We want a homogeneous development and execution environment.

CPUs have had to give up the fiction of a single thread of execution, which is good, because we don’t have to deal with the thermal issues of 10 GHz out-of-order execution, but which is bad, because we have to rewrite all of our code to be multi-threaded.

GPUs got a jump start on multi-threaded execution of a narrow set of algorithms, which meant lots and lots of relatively slow units making graphics go fast. CPUs now are going to have lots and lots of somewhat faster units making everything go fast including graphics, if they can get people to rewrite their software. GPUs are looking around for other software they can get to run on their cores, and eventually do all the ugly things CPUs have to do in order to run operating systems, spreadsheets, and the like.

Whatever the transistors you have in your future box, you want them to all be running the same instruction set with unified memory with great compilers and debuggers without some API sitting in the middle of your threads (e.g. physics API or DirectX). If you have enough money to put a second big chunk of transistors in your box, you want them to be the same as the first – not introduce some heterogeneous set of headaches for software developers.

I hope this makes sense.

Make of that what you will, I'll reply and update if I get another response.
 
You stop the FUD-mark in a pipe a smoke it, in case I wasn't clear, I am speaking about GAMEPLAY, not FUD-mark..the picture changes when doing real-world gameplay...

3DMark Vantage is a synthetic test, showing the maximum potential for physics processing.
If you were to use a GPU entirely as a physics processor (which theoretically you could, with a multi-GPU setup), this is the performance it would deliver.

In REAL gaming (like UT3) the PPU is still king...not to bad for a 2006 130nm tech PPU wit a TDP of 28 Watt...GPU's look so like bruteforce spaceheaters compared....that is not impressive...that is sad.

So far we've only seen it tested with a 9800GTX.
We don't know how much faster the PPU can get. It can keep up with the GPU at 60 fps. But perhaps it starts to limit the GPU if you reach 70 fps (which I wouldn't find too surprising, since the PPU delivers about twice the performance of a quadcore in 3DMark Vantage).
So the tests will have to be repeated with faster GPUs (or lower resolutions) to find where the limits of the PPU are. And then I'd like to see where the limits of a GTX280 are.
 
Make of that what you will, I'll reply and update if I get another response.

I think he's just expressing his utopia of a single generic processing unit. Sure that will be easy to develop for, but that's not going to happen in the next few decades.
We have to work with reality, and the reality is that we have custom-designed processors for various tasks, such as audio, video, networking etc, because they are far more efficient and cost-effective than using the CPU for everything.
Just look at the built-in AES encryption on the VIA CPUs for example. Just a simple and cheap add-on, but faster than the fastest quadcore. If you want everything in generic x86-cores, you'd have to throw in far more transistors.

If you interpret what he says literally, he'll be replacing Direct3D with his own software renderer, and replace Havok with his own physics library. But he hasn't done that with any of his previous products, and I doubt he'll do that with his next product.

Basically he's being diplomatic and not saying what he's actually going to do in the near future.
 
3DMark Vantage is a synthetic test, showing the maximum potential for physics processing.
If you were to use a GPU entirely as a physics processor (which theoretically you could, with a multi-GPU setup), this is the performance it would deliver.

Okay, now you go on my ignore...anyone promoting FUD-mark as a performance-messaure-tool is lost.
 
Okay, now you go on my ignore...anyone promoting FUD-mark as a performance-messaure-tool is lost.

So just because it's called 3DMark, performing a bunch of physics operations using the standard PhysX API and measuring how many operations per second you perform, is suddenly no longer a performance-measurement?

Wow, with logic like that, who needs any kind of measurements anyway :)

The article linked to by Cybereality on benchmarkreviews.com shows pretty much the same anyway, GPU being about 10-11x faster with physics (even though they used the PhysX SDK samples rather than 3DMark Vantage).
 
The reality is that the developers decide what is best for them, and use that.

It certainly doesn't sound to me as if he's enthusiastic about expanding past the processor if it can be avoided, everything else he's said in interviews has been along those lines.

If people like Gabe can't be convinced to send the direction of their engine that way then physics on the GPU may have some troubled times ahead, considering the playerbase valve has across all its titles and the kind of money they make. They're withotu a doubt one of the biggest players in the industry.
 
If people like Gabe can't be convinced to send the direction of their engine that way then physics on the GPU may have some troubled times ahead, considering the playerbase valve has across all its titles and the kind of money they make. They're withotu a doubt one of the biggest players in the industry.

"People like Gabe" being the key there.
Are all developers "people like Gabe"? I doubt it. Gabe showed no interest in D3D10, even though quite a few games have included support for it by now, and in fact D3D10.1 is supported by some aswell.
UnrealEngine 3.0 already supports PPU/GPU accelerated physics, and that's not exactly a small player either.

Gabe Newell, Tim Sweeney, John Carmack, Cevat Yerli... none of them seem to agree on many things. I think you shouldn't really see the game industry has having one common view or goal.
 
You are working from the assumption that framerate is the primary goal of a game.
This is not the case. For years, developers have been adding extra features that impact framerate, but improve the overal appearance and/or gameplay.
Physics is just another of those features.
They're just taking advantage of the added power of new hardware.

No, I'm working from the assumption that we dont have any of these features in todays games (in any significant way), and that you have to sell the idea to developers, that inlcudes the advantages over our current system and the disadvantages.

People here are assuming that somehow everyone is FORCED into using cloth physics, and since the GPU is better at that, so thats the best and theres no alternative. THIS IS WRONG.

The developers have to decide if having 1-2 sheets of cloth blowing in the wind slightly more realisticly is worth the rendering time its going to cost to do that, which at the moment looks to be in the order of 25% in UT3

You're assuming that games are still singlecore. Most current-gen game engines are already optimized for multicore, so this assumption is wrong. Aside from that, at least Havok and PhysX, the two main physics APIs, are optimized for multicore. So even if the rest of your game is not multicore, the physics part will be, if you use either library. So your argument is false. The only untapped potential is because games are 'neutered' to run well on low-end CPUs.

The majorety of games are still single core, there is a handful of games which exceed that to dual core, and theres about 2 games which use quad core, none of them are especially elogant at the moment.

Go play some actualy games and check your core usage, you'll only see quake4, UT3 and crysis attempt to use all 4 cores and thats about it, some games will attempt to use 2 cores and not do it especially well, and most games will only use 1 core

If you run the CPU benchmarks included with Crysis on a quadcore at very high physics, you'll see that the CPU is absolutely not spared at all. You're still CPU-limited once that rocket launcher starts blasting.
Same with 3DMark Vantage, it dumps a shitload of physics on all 4 cores... It even adjusts the number of objects in the test to the number of cores you have.
So the problem isn't at all that we can't use quadcores or anything.

Crysis is 1 game...out of how many? Besides the number of physics objects in Crysis is quite impressive, you can drive a jeep into the middle of a hut crammed with objects and shove some barrels in there as well and blow the lot and send everything flying.

The question is what happens when we take the physics load from Crysis and put that onto the GPU, instead of neatly using the additional power of the spare cores it suddenly has to use a GPU which is already at 100% capacity straining to render the game as it is, that results in frame rate loss.

Why the hell does anyone care about 3DMark Vantage, thats a piece of shit like all the other 3Dmarks, that doesn't tell us anything helpful about how this works in real world scenarios. Anyone with any brains stopped relying on 3Dmark for anything outside of stability testing, a long time a go.

It does, because the cost in framerate is smaller on the GPU than it is on the CPU. Aside from that, the CPU is freed up for other tasks, that the GPU wouldn't be able to do (efficiently). So developers can add more features to the rest of the game.
It's all about getting the most from the hardware. Arguing against physics acceleration is about as useless as arguing against graphics acceleration.

Sacraficing huge chunks of the frame rate and thus having to turn down all your settings isnt using the hardware better, it's swapping around one graphical feature for another, the GPU in modern games is always stressed 100% you cannot add more workload and maintain acceptable frame rates, thats what this whole thread is about, re-read the title "One GPU can NOT handle physx + video: benchmarks"

No matter how you spin it, this remains true.
 
The majorety of games are still single core, there is a handful of games which exceed that to dual core, and theres about 2 games which use quad core, none of them are especially elogant at the moment.

Go play some actualy games and check your core usage, you'll only see quake4, UT3 and crysis attempt to use all 4 cores and thats about it, some games will attempt to use 2 cores and not do it especially well, and most games will only use 1 core

You forgot about WiC and SC...but then again those games don't support your conclusion...
 
No, I'm working from the assumption that we dont have any of these features in todays games (in any significant way), and that you have to sell the idea to developers, that inlcudes the advantages over our current system and the disadvantages.

People here are assuming that somehow everyone is FORCED into using cloth physics, and since the GPU is better at that, so thats the best and theres no alternative. THIS IS WRONG.

The developers have to decide if having 1-2 sheets of cloth blowing in the wind slightly more realisticly is worth the rendering time its going to cost to do that, which at the moment looks to be in the order of 25% in UT3

I'd say it's inevitable. I can name many things that at one point were incredibly costly (think about per-pixel lighting, or realtime shadows), but as hardware became more powerful, it became a standard feature of games. One could argue that per-pixel lighting and realtime shadows don't really add much to the overall gameplay, but we all know that's not how the story went.

At some point, one developer will decide that the time is right for cloth-effects on all its characters and such, and then all games lacking realistic cloth-effects will look fake and outdated, so other developers will add it aswell.

The majorety of games are still single core, there is a handful of games which exceed that to dual core, and theres about 2 games which use quad core, none of them are especially elogant at the moment.

Go play some actualy games and check your core usage, you'll only see quake4, UT3 and crysis attempt to use all 4 cores and thats about it, some games will attempt to use 2 cores and not do it especially well, and most games will only use 1 core

I've already explained how this has nothing to do with how efficient the game really is. I'm not going to repeat myself.

Crysis is 1 game...out of how many?

Crysis is the first of a new generation of games.

Besides the number of physics objects in Crysis is quite impressive, you can drive a jeep into the middle of a hut crammed with objects and shove some barrels in there as well and blow the lot and send everything flying.

Yes, but there's still so much that the physics in Crysis DON'T do. As nice as Crysis already is, there's still plenty of room for improvement.

The question is what happens when we take the physics load from Crysis and put that onto the GPU, instead of neatly using the additional power of the spare cores it suddenly has to use a GPU which is already at 100% capacity straining to render the game as it is, that results in frame rate loss.

Well, the load of the Crysis physics is less than that of the UT3 PhysX levels.
Also, the GPU isn't at 100% capacity. Some parts are bottlenecked, so it cannot render more frames, but that doesn't mean its entire processing capacity is utilized.
I repeat for the umpteenth time, these are not the same things!
If pixel shaders have to wait because a shadow map has to be sampled, it cannot go faster, but the pixel shaders are sitting idle.

Why the hell does anyone care about 3DMark Vantage, thats a piece of shit like all the other 3Dmarks, that doesn't tell us anything helpful about how this works in real world scenarios. Anyone with any brains stopped relying on 3Dmark for anything outside of stability testing, a long time a go.

The physics test doesn't even TRY to be a real-world test. It just tries to determine how quickly your system can run physics. It's a purely synthetic test, just like the fillrate test for example.
Bash 3DMark all you want for it not being an indicator for how fast <insert your favourite game here> runs, but trying to deny the usefulness of a synthetic physics test is rather far-fetched. Synthetic tests aren't real-world scenarios by definition.

Sacraficing huge chunks of the frame rate and thus having to turn down all your settings isnt using the hardware better, it's swapping around one graphical feature for another, the GPU in modern games is always stressed 100% you cannot add more workload and maintain acceptable frame rates, thats what this whole thread is about, re-read the title "One GPU can NOT handle physx + video: benchmarks"

No matter how you spin it, this remains true.

Actually, you're wrong. However, I have lost the interest in trying to explain it yet again.
I can only suggest you read some articles on how G80 schedules threads, instructions etc. You'll see that the GPU still holds untapped potential.
 
Wow...sorry 4-5 games instead of 2-3, yeah that makes your liabry of games absolutely staggering now, i've completely reversed my opinion :rolleyes:
 
You can throw names all you want. But this 'anonymous forum poster' is a professional software developer, who has hands-on experience with both NovodeX/PhysX and Cuda code, and you don't.
Now if you want to claim that the way the G80 architecture executes its threads has analogy with HyperThreading whatsoever, that only further illustrates your lack of insight in the matter.
You probably wouldn't even know where to begin writing anything with PhysX or Cuda. Let alone that you actually set up any test cases and analyzed performance characteristics, which this 'anonymous forum poster' has.
So I'm talking from first-hand experience, something you could never do.

You got me there. I am sure you have a lot more exposure to the world of PhysX locked in your cube coding C. God knows you won't tell us which one you work for, lest we all find out that Hello Kitty PhysX Edition is the blockbuster that cometh. :rolleyes: Have a good day Scali2, if that is your real name.
 
I doubt that this "professsional software developer" knows much about market research, profitability, cost management and etc. because eventhough the performance benefit with the right GPU could be huge, that alone won't be enough to push developers to adopt PhysX soon.

The only thing he knows is what the NVIDIA / PhysX evangelist has been pushing down his throat while throwing marketing money at his company. But I am sure he can't talk about that.
 
You stop the FUD-mark in a pipe a smoke it, in case I wasn't clear, I am speaking about GAMEPLAY, not FUD-mark..the picture changes when doing real-world gameplay...

Show me one instance off this, and I will consider your argument...untill then, it in the same box a FUD-mark....the never-never-land...as test whow it the GPU becommming bottleneck when there is to much particle physics going on.

Oh, really...based on what?
FUD-mark?

In REAL gaming (like UT3) the PPU is still king...not to bad for a 2006 130nm tech PPU wit a TDP of 28 Watt...GPU's look so like bruteforce spaceheaters compared....that is not impressive...that is sad.

he has nothing more to bank on as he obviously has not real world experience to point to besides research done on UT3 by another site.
 
You are working from the assumption that framerate is the primary goal of a game.
This is not the case. For years, developers have been adding extra features that impact framerate, but improve the overal appearance and/or gameplay.
Physics is just another of those features.
They're just taking advantage of the added power of new hardware.



You're assuming that games are still singlecore. Most current-gen game engines are already optimized for multicore, so this assumption is wrong. Aside from that, at least Havok and PhysX, the two main physics APIs, are optimized for multicore. So even if the rest of your game is not multicore, the physics part will be, if you use either library. So your argument is false. The only untapped potential is because games are 'neutered' to run well on low-end CPUs.
If you run the CPU benchmarks included with Crysis on a quadcore at very high physics, you'll see that the CPU is absolutely not spared at all. You're still CPU-limited once that rocket launcher starts blasting.
Same with 3DMark Vantage, it dumps a shitload of physics on all 4 cores... It even adjusts the number of objects in the test to the number of cores you have.
So the problem isn't at all that we can't use quadcores or anything.



It does, because the cost in framerate is smaller on the GPU than it is on the CPU. Aside from that, the CPU is freed up for other tasks, that the GPU wouldn't be able to do (efficiently). So developers can add more features to the rest of the game.
It's all about getting the most from the hardware. Arguing against physics acceleration is about as useless as arguing against graphics acceleration.

Install base and profitability are the game devs priorities unless they have big marketing money behind them.

And if you did not note, the guy you keep talking to like an idiot was THE person that brought gameplay evaluations to lite in the world. So I do not think it is all about framerate, but you refuse to see that PhysX is far from free.
 
Okay, now you go on my ignore...anyone promoting FUD-mark as a performance-messaure-tool is lost.

He has nothing other to base his arguments on. That and UT3, woohoo.
 
The reality is that the developers decide what is best for them, and use that.

It certainly doesn't sound to me as if he's enthusiastic about expanding past the processor if it can be avoided, everything else he's said in interviews has been along those lines.

If people like Gabe can't be convinced to send the direction of their engine that way then physics on the GPU may have some troubled times ahead, considering the playerbase valve has across all its titles and the kind of money they make. They're withotu a doubt one of the biggest players in the industry.

One thing that Valve has done better than any other game dev, is that it has continually developed products that pushed the envelope that still were able to be sold to an huge install base that were still able to play the game the and have great fun with it.

Gabe is not too much worried about it because Valve has first and foremost in my opinion been concerned with bring great gameplay to market, rather than the next graphics block buster. Compare HL2:EP2 sales and Crysis. Who won out there?

Now look at what Valve has done with Physics in their games. Valve is the only company I know that has won Game of the Year awards with two games based SOLELY on in-game physics. From Portal to Peggle.

No matter what your stance is on this argument, you cannot argue that "Valve is doing it wrong."
 
Crysis is 1 game...out of how many? Besides the number of physics objects in Crysis is quite impressive, you can drive a jeep into the middle of a hut crammed with objects and shove some barrels in there as well and blow the lot and send everything flying.

The question is what happens when we take the physics load from Crysis and put that onto the GPU, instead of neatly using the additional power of the spare cores it suddenly has to use a GPU which is already at 100% capacity straining to render the game as it is, that results in frame rate loss.


Sacraficing huge chunks of the frame rate and thus having to turn down all your settings isnt using the hardware better, it's swapping around one graphical feature for another, the GPU in modern games is always stressed 100% you cannot add more workload and maintain acceptable frame rates, thats what this whole thread is about, re-read the title "One GPU can NOT handle physx + video: benchmarks"

No matter how you spin it, this remains true.

We have a winner that sees the truth about GPU physics.

A question worth thinking about is why did ATI walk away from GPU physics the moment it had a CPU to rely on? You can come to several different answers that you think are true, but the point is to think about the possible variables behind that call.
 
Having PhysX effects like clothes and particles are nice but it is not neccessary. I would rather have a game which the developers have spent more time into the gameplay than having a game with a lot of physics but using the same old gameplay.

Wii has high sales not because of the graphics but because it is fun to play. MGS4 is also a great game which the developers have paid a lot of attention into the small details that make the gameplay really interesting eventhough the actual gameplay is quite short but it doesn't matter because the game will make you want to play it again with another method.
 
Actually, I'm not even in the gaming industry, and therefore not under the attention of the nVidia marketing forces.
We're mainly interested in Cuda for accelerating our scientific data processing.
Since we combine this with data visualization, we're quite aware of how graphics and Cuda can coexist on a single system (even with relatively simple Dell laptops with Quadro chips that we use around the office here).
I've only toyed a bit with PhysX to see how fast and accurate it was, and if it could be of any use in a more 'serious' environment than gaming, and because I like to write demo/game engines in my spare time.

Now I don't see any use in telling the world what company I work for. Firstly, I prefer a bit of privacy/anonymity, and secondly I don't speak for my company, but purely for myself, so there is no relevance in mentioning it anyway.
However, if you really want to know, you could always check your logs for my IP. I expect you to respect my privacy however.

Thanks for that explanation. So you have no experience with PhysX and gaming? Given that I you do or you don't, please share some of these demo and game engines you write in your spare time. Given your arguments here, I still think you are just some guy locked in a cube writing C looking at UT3 benchmarks to somehow prop up your thoughts. And given how much time you spend posting and reading here, I would suggest your workload is not that intense. I would highly suggest that I have installed and tested more PhysX cards than you have ever had access to. So your suggestions that I am no way experienced enough to have an opinion is laughable at best given your situation.

How many major AIBs came to you to talk to you about AIC physics before they were close to reality? And by the way, I called it exactly, they did not listen to me and are left with 10s of thousands cards they will never sell. Some say hundreds of thousands in dead inventory.

I respect your arguments on PhysX and understand your viewpoint, but you do not see the forest for the trees. Unless it is a product that game developers can use without narrowing their install base and without major impacting time to market, then it is technology that is not good for them and will not make it to the user. And if more than a few AAA titles are not going to push it out in a successive fashion, it is good as dead. UT3 is the biggest title PhysX has been attached to and I would suggest that it had no relevance in the games sales or hardware sales. Hopefully others will come.

I have seen for years now exactly what PhysX can do in game when it comes to bringing forth another layer that will further immerse the gamer in content. No doubt it can do some very impressive things. It has just yet to be seen if we can do it all on one GPU. I personally think that more gamers would be interested in buying an add-in card at this point. Any number of companies should be able to produce these very cheaply now if the GPU load is as low as NVIDIA says it is. That way you would be able to use your GPU-bottled necked game for graphics on the GPU.

I think in the age of expanding display size, more gamers find it important to have more pixels in front of them and I am not sure the GPU is in a position to be leverage to give up a good chunk of its power for PhysX. Especially when we have multicore CPUs lying around doing "nothing." I suggest most gamedevs will look at that and decide where they want they power to come from and what they want to spend their time coding.

That is all I have to say on the subject. Nothing written here by anyone has changed my mind. Same as my meeting with Manju a few weeks back when he personally showed me all of his PhysX demos, so of which were incredible. If you ever get a chance to see AI developed through NVIDIA lobster/slug/sea anemone demo, you will not want to miss it. The most impressive thing I have seen done on GPU physics yet in terms of gaming.
 
Havok FX was just an ill-conceived attempt at GPU-based physics.
However, AMD itself announced that they're back on researching GPU-assisted physics with Havok:
http://www.amd.com/us-en/Corporate/VirtualPressRoom/0,,51_104_543~126548,00.html



So apparently AMD/Havok want it both ways: core game physics on CPU and other types of physics simulation on the GPU.

Now we can all speculate on why this is ofcourse... Perhaps nVidia chose to do everything on the GPU because they don't own a CPU themselves, and try to steal Intel/AMD's thunder... Or perhaps AMD's GPUs aren't as good at physics as the nVidia GPUs are, and for AMD it's not a viable option to go all GPU.
Only time will tell what AMD/Havok will turn up with.
And then ofcourse there's Intel, that will most probably (ab)use Larrabee for physics. I really don't think Intel would have bought Havok if they were only interested in CPU physics. After all, Havok already had full CPU support. What purpose does it serve Intel to own that company?

Drum role please...GPU physics will be here undoubtedly. GPU PhysX? I think that is another story. AMD and Intel are hell bent enough on NVIDIA destruction that they will partner and work around anything with NVIDIA branding on it.
 
I'd say it's inevitable. I can name many things that at one point were incredibly costly (think about per-pixel lighting, or realtime shadows), but as hardware became more powerful, it became a standard feature of games. One could argue that per-pixel lighting and realtime shadows don't really add much to the overall gameplay, but we all know that's not how the story went.

At some point, one developer will decide that the time is right for cloth-effects on all its characters and such, and then all games lacking realistic cloth-effects will look fake and outdated, so other developers will add it aswell.

I can name many things that also DIDNT make it, I think you know that list as well as I do so I wont bother with all the failed attempts throughout history, I will point out that the PPU was one of them however...

I've already explained how this has nothing to do with how efficient the game really is. I'm not going to repeat myself.

Its not down to efficency, some games were just never deisgned to use all the cores to start with, they're perefectly efficient for what they are, it just means theres untapped power to get more physics in there like 2-3x what a single core.

Crysis is the first of a new generation of games.

Exactly, and look what it provides when it uses quad core properly, I'm not about to start downplaying how impressive Crysis is, in comparison to other games out today it's got very decent level of physics which most importantly are effective in the game world, so you can use them as cover, and AI can tell where a tree has fallen and use it as cover.

Yes, but there's still so much that the physics in Crysis DON'T do. As nice as Crysis already is, there's still plenty of room for improvement.

There's always room for improvement, i could say that about the PPU, it's about what is fitting for the hardware at the time, even Crysis was ahead of its time by a good year or so.

Well, the load of the Crysis physics is less than that of the UT3 PhysX levels.
Also, the GPU isn't at 100% capacity. Some parts are bottlenecked, so it cannot render more frames, but that doesn't mean its entire processing capacity is utilized.
I repeat for the umpteenth time, these are not the same things!
If pixel shaders have to wait because a shadow map has to be sampled, it cannot go faster, but the pixel shaders are sitting idle.

Yes Crysis physics are less, but this is a perfectly reasonably alternative to a developer like Valve for example, they have a choice to either persue something like liquid and cloth physics and then decide what is best to do them on (and we both agree gpu) or aim for lower levels of physics and do that for free (no performance penalty) by using multicore well.

The physics test doesn't even TRY to be a real-world test. It just tries to determine how quickly your system can run physics. It's a purely synthetic test, just like the fillrate test for example.
Bash 3DMark all you want for it not being an indicator for how fast <insert your favourite game here> runs, but trying to deny the usefulness of a synthetic physics test is rather far-fetched. Synthetic tests aren't real-world scenarios by definition.

Yet you use it as an example of somthing where people are only interested in the real world application?

Actually, you're wrong. However, I have lost the interest in trying to explain it yet again.
I can only suggest you read some articles on how G80 schedules threads, instructions etc. You'll see that the GPU still holds untapped potential.

I'm not wrong, the frame rates in the OP clearly show the FPS difference.

You maybe correct that there are some unused cycles on the GPU to do something like physics calculations but certainly this is not a reliable way of doing things, when exactly it has idle time to spare and what cards actually have idle time to begin with and how much they have all vary incredibly. UT3s frame rate drops show this perfectly in a real world scenario.

I honestly do understand that its not that simple and that video cards are complex and may have spare processing time spare while other parts of the card are bottlenecked but this is fractional at best, we're talking a few percent at most which isnt a significant saving...GPUs are already fairly efficient.
 
<3 Kyle :)

I'm only arguing the technical merits of the G80 architecture, Cuda and the PhysX implementation of it. I don't wish to engage in matters of politics or marketing, as I said before.

As I said early on in my argument, the real world application of this stuff is what is important, you can argue the benchmark numbers until you're blue in the face, but unless it makes sense from a developers point of view then it's never going mainstream.

Politics and marketing will make or break something like this, especially if it's a hard sell to begin with.

I dont think anyone disagrees with the GPU being more efficient or really disagrees with a lot of the technical arguments you have, just that it doesn't relate directly to anything in real world scenarios.
 
I dont think anyone disagrees with the GPU being more efficient or really disagrees with a lot of the technical arguments you have, just that it doesn't relate directly to anything in real world scenarios.

Well, I was mostly responding to this thread because the OP claims otherwise.
It tries to claim that GPUs cannot do PhysX... That is what I was arguing against.

Arguing that game developers might not take advantage of GPU PhysX is an entirely different topic altogether, and one I would likely not even respond to. Only the game developers themselves know what they are or aren't going to do with GPU PhysX, all we can do is wait and see. Some people are of the opinion that they are, others are of the opinion that they aren't. I don't see much of a point in arguing over that.
 
Well, I was mostly responding to this thread because the OP claims otherwise.
It tries to claim that GPUs cannot do PhysX... That is what I was arguing against.

Arguing that game developers might not take advantage of GPU PhysX is an entirely different topic altogether, and one I would likely not even respond to. Only the game developers themselves know what they are or aren't going to do with GPU PhysX, all we can do is wait and see. Some people are of the opinion that they are, others are of the opinion that they aren't. I don't see much of a point in arguing over that.

No, you were directly responding to many others besides the OP.

The OP NEVER claimed that GPUs cannot do PhysX.

And you obviously want to argue over points other than what you state above since you keep responding to statements on that very topic.

I think you are confused.
 
As said before, this is not the case. GPUs are so much faster that CPUs simply cannot compete at all with physics.
What benchmarks have shown so far is that you need roughly 10% of the processing resources of a 9800GTX to match a high-end quadcore.
Ergo, a quadcore is 100% loaded with physics is about equivalent to a 10% loss in framerate if the physics workload was done by the GPU.
Extrapolating from there, if you were to load the quadcore with twice the workload ('200%'), you get a 50% drop in framerate, however for the GPU that would mean 20%.
So GPUs scale much better. The heavier the physics workload is, the more sense it makes to offload it to the GPU, as counter-intuitive it might sound.

I understand completely what you are saying. However, though this technique and technology may work on a 9800GTX, what about lower-end graphics cards such as a 2400XT or an 8600GT. Most people are gaming with these types of cards. The people who play with 9800GTX's is a much smaller portion of gamers.

So if the 9800GTX is faster than a high-end quad core CPU, then what about the low-end cards? Would this result in a 40% drop in framerate or more by using a 2400XT or an 8600GT.

This is my only problem with it. I understand that it can work with high-end graphics cards, but what about the low-end cards?

Also, don't tell me that people with low-end cards are screwed. I explained why in my last post that it would be bad for the developers and publishers to isolate the low-end graphics crowd and would thus cause horrible sales.
 
I can name many things that also DIDNT make it, I think you know that list as well as I do so I wont bother with all the failed attempts throughout history, I will point out that the PPU was one of them however...

The two main reasons why the PPU failed:
1) It required users to buy an extra piece of hardware in their system, which was of little use outside games that supported it.
2) It required applications to support it, and a 'killer app' never emerged.

nVidia has now eliminated 1).
A lot of people already bought the hardware, before even knowing it would support PhysX.
It's now up to nVidia to eliminate 2).

Its not down to efficency, some games were just never deisgned to use all the cores to start with, they're perefectly efficient for what they are, it just means theres untapped power to get more physics in there like 2-3x what a single core.

Like I already said, both Havok and PhysX are already optimized for multicore.
No matter how single-threaded the rest of your game is, if you use either of these libraries, your physics will use all available CPU cores.
So for physics, the multi-core argument simply doesn't hold anymore, it's solved.

Exactly, and look what it provides when it uses quad core properly, I'm not about to start downplaying how impressive Crysis is, in comparison to other games out today it's got very decent level of physics which most importantly are effective in the game world, so you can use them as cover, and AI can tell where a tree has fallen and use it as cover.

From what I've seen, the performance on a quadcore is virtually identical to a dualcore.
It uses the extra cores, but it doesn't gain a lot of performance, because the engine just doesn't scale that well. Most multicore-engines don't... It depends on the game. RTS tends to scale better than FPS, because it relies more on AI and things.

Yes Crysis physics are less, but this is a perfectly reasonably alternative to a developer like Valve for example, they have a choice to either persue something like liquid and cloth physics and then decide what is best to do them on (and we both agree gpu) or aim for lower levels of physics and do that for free (no performance penalty) by using multicore well.

I think it all depends on what the average system will look like.
Today, if you buy a $200 CPU and a $200 GPU, you're likely to have a fast dualcore and a 9800GTX or something.
In which case, GPU physics will be more interesting.
If however, you choose some kind of office PC, eg a $1000 quadcore CPU and onboard graphics, then CPU physics will be the way forward.

Yet you use it as an example of somthing where people are only interested in the real world application?

No, I compared the potential of a GPU doing only physics (well not entirely, Vantage still does some graphics), to the potential of a PPU, to illustrate what things could look like if you were to use two GPUs, one for physics, one for graphics.
That dedicated GPU would then be considerably faster than the PPU, and as such, wouldn't limit the other GPU as quickly as the PPU might, thus yielding higher in-game framerates. That was the point I was making.
I don't see anything wrong with using 3DMark Vantage in this instance. In fact, since Vantage still renders graphics during that test, it is actually slightly favourable to the PPU.

I'm not wrong, the frame rates in the OP clearly show the FPS difference.

The point I'm making is that some of the physics are 'free' because they run in 'idle time' of the GPU.
You can more or less see this by comparing the hit in framerate between the two resolutions. The hit at the high resolution is larger, because at a high resolution the GPU has less 'idle time' (overly simplified version: polygons exist of more pixels, making better use of parallelism during rasterizing).

I honestly do understand that its not that simple and that video cards are complex and may have spare processing time spare while other parts of the card are bottlenecked but this is fractional at best, we're talking a few percent at most which isnt a significant saving...GPUs are already fairly efficient.

nVidia GPUs are estimated to have between 70-90% efficiency. This leaves 10-30% idle time for physics. This gets you a long way, although obviously the UT3 levels push it further than that.
The physics load in your average game (eg the normal UT3 levels) is probably 'free' on GPU PhysX.
And when you have a GTX280 instead of a 9800GTX, you can stand to lose a few more frames anyway.
Since hardly anyone owns a PPU, and you can hardly buy them anymore, I think GPU PhysX is an excellent substitute.
 
I understand completely what you are saying. However, though this technique and technology may work on a 9800GTX, what about lower-end graphics cards such as a 2400XT or an 8600GT. Most people are gaming with these types of cards. The people who play with 9800GTX's is a much smaller portion of gamers.

So if the 9800GTX is faster than a high-end quad core CPU, then what about the low-end cards? Would this result in a 40% drop in framerate or more by using a 2400XT or an 8600GT.

This is my only problem with it. I understand that it can work with high-end graphics cards, but what about the low-end cards?

Also, don't tell me that people with low-end cards are screwed. I explained why in my last post that it would be bad for the developers and publishers to isolate the low-end graphics crowd and would thus cause horrible sales.

If your GPU cannot run physics, you simply don't enable it. If your CPU cannot run physics either, well yes, then you are screwed, this has always been the case.
I don't see your point really. If they don't have CPUs that are fast enough, they can't play the games either. It's just a fact of life, games get more demanding every year.
Most games don't run, or don't run well on my 5-year old system either.
You have to upgrade at some point.

Besides, the high-end cards of today are the low-end cards of tomorrow. At some point even integrated video will be faster than a 9800GTX.
 
PhysX+video, obviously.

No, this is what the OP said.

So, this confirms what I and several other people have suspected: when the GPU is under heavy load, it can not handle 3D and physx at the same time.

(And this is why we supply you with a quote button. Because what you are saying he said, and what he said are two very different things. Don't put words in peoples' mouths when there is no need.)

And he is 100% correct. When a GPU has a full traditional graphic load, it is not going to handle PhysX physics well, no matter how much efficiency you "hyperthread" into it.

I have stood in a room with Manju many times, the godfather of PhysX, and made the same argument, and he has NEVER given me any reason to think otherwise than what has been posted here.

So apparently some Cuda coder that does not develope games locked in a cube in Europe knows more about PhysX than I do or the person that "invented" it. You are simply misleading in many of the statements you have made here in this thread or have quite simply do not have the English skills to properly convey what you wanted to say. Either way, it is done. Time will tell what is to become of PhysX.
 
Status
Not open for further replies.
Back
Top