PhysX only using one cpu core?

[Benetanegia]

Point A) OMG trash on the ground, that is the best part about physX!

That is showing the power of the GPU running physics. And you have never seen anything like that running in the CPU, otherwise link. What you (all I must say) fail to see, is that "anyone" can create almost any effect, it's simple maths after all. The key is in how many calculations you can make. The way of making smoke or cloth look better is by having more particles or nodes.

Point B) all of the games that use PhysX suck the big one, do I really need to post links? (and No UTIII is not a physx game, it has one level that uses phyx and you need to DL it separately)

UT3 does use PhysX in the regular game. It has one level that uses GPU accelerated PhysX and in reality they are 3.

And yeah all of them suck? I don't know your tastes, but Mass Effect sucks? Gears Of War sucks? Age of Empires III sucks? Brothers In Arm: HH sucks? City of Villains sucks? Tom Clancy games sucks? Virtua Tennis 3 sucks? Mirror's Edge sucks?

I dunno, the buyers seem to differ with you.

Point C) your completely missing the point that all of the magic wonderful amazing GPU PHYSX CRAP you are promoting can easily be run on a CPU, now, today using a proper cpu physics engine. MOST IMPORTANTLY: Without a hit to FPS in the game!

And which is that proper magical physics engine that is nowhere to be found? Why is Havok so stupid that they don't make it for once? No, the reality is that a CPU can't do the ammout of physics that are done in Nvidia cards right now and that's why Havok, AMD, Intel and M$ are putting so much effort and money into implementing their own GPU accelerated physics, while brainwashing all of you until they have their own competing product. Time to a reality check.

now refute the points WITHOUT changing the subject or GTFO[/QUOTE]

 

[SNiiPE_DoGG]

the Confirmation Bias is strong with this one....

I wish I could find the demo of i7 running physics on Havoc with thousands of entities.... google is not turning it up though - was in the news like ~4 months ago IIRC

 

[Benetanegia]

the Confirmation Bias is strong with this one....

I wish I could find the demo of i7 running physics on Havoc with thousands of entities.... google is not turning it up though - was in the news like ~4 months ago IIRC

yeah I wish you could because no combination of physics+core i7+havok gives any result. What it does appear a lot is the OpenCL accelerated Havok running in AMD cards, which showed many entities indeed and was about 4 months ago IIRC. I think you are confused with that. I read the tech press everyday and I have never heard of that, plus there's nothing in google about that, so IMO it's fake until I see it. I'll continue sarching.

And I'm not biased at all. I've tried many physics demos from various companies and open source and I have a very decent knowlegde about what my CPU can handle. A core i7 can handle twice. Until you show me something, your claim is a lie.

 

[KainXS]

Good Lord man, you have been here since early this morining defending PhysX, give up on it, PhysX is and has always been coded like sh*t for cpu's, . . . . why, because PhysX was originally one thing, ageia's cash cow, in software mode it dosen't do anything but waste cpu cycles in my opinion, I look at on one hand Havok, which runs very very good on cpu's it may have a few less features then but its worth it in the end because its universally compatible and you don't need a super fast CPU or a certain GPU or card to use it, for devs theres not much of a risk. then I look at PhysX, I remember back PhysX came out, the GRAW tests with PhysX, they were absolutely terrible, its improved a bit since then but not as much as it needs to to really take off.

Now with the possiblilty that DX11 could be a decent alternative for free with no strings attached, I say let PhysX and Havok die, for a while now, Havok has been on its deathbed, and I think PhysX its on its deathbed also since you can't use gpu physX at all if you have an ATI card now, I say go DX11.

I'm out.

 

[Benetanegia]

Good Lord man, you have been here since early this morining defending PhysX, give up on it, PhysX is and has always been coded like sh*t for cpu's, . . . . why, because PhysX was originally one thing, ageia's cash cow, in software mode it

Ageia has always have a CPU path and it ran very well in comparison to other CPU physics engines at the time including Havok. Ghost recon was poorly coded, period. It has nothing to do with the API. Havok has been used in better games in the past, that's the difference. Like HL2, it´s Valve's merit and not Havok's, but havok has always taken the merit for that.

 

[SNiiPE_DoGG]

Found it! wasnt havok, I was mistaken, but on the CPU nonetheless

http://www.viddler.com/explore/HardOCP/videos/36/

 

[FordGT90Concept]

Now with the possiblilty that DX11 could be a decent alternative for free with no strings attached, I say let PhysX and Havok die, for a while now, Havok has been on its deathbed, and I think PhysX its on its deathbed also since you can't use gpu physX at all if you have an ATI card now, I say go DX11.

DirectX 11 has no physics APIs, only the capability to run physics code on any DX11 GPU using the compute shaders. The question is: Who is going to author the physics API for DX11? Once you have an answer to that, I think you'll have an answer for the most dominent phyisics API 10 years from now.


What has worked in the past and remains very popular is needs-based physics. That is, you code it in house and make it as accurate as deemed necessary. An open, free-to-everyone, physics standard isn't very likely.

 

[Benetanegia]

Found it! wasnt havok, I was mistaken, but on the CPU nonetheless

http://www.viddler.com/explore/HardOCP/videos/36/

Ok, those are some nice number of objects there for a processor. But you do realize that the one on the GPU has much much more processing going on right right? That's point number one. I don't remember if they say there, but in this one you posted is 1500 Boxes and 200 ragdolls are maxing out the i7. In one of the links I posted a car trail is being simulated with 250.000 particles, that's a lot more (like ten times more despite not being solids and ragdoll), and that ratio will never change. A GPU can still do much more things.

Point number two, I never said that a CPU like i7 using all of it's cores can't do a high ammount of physics, I said that the CPU that most people have won't be able to come even close (i7 is 5x faster than the average dual core, not to mention console CPUs and single cores), I ve been saying that constantly. Developers have to develop for the lower end if they want to sell. What's the percentage of people that have a core i7? Not even the 2%. It makes no sense to develop a game with that base in mind. And unless you scatter many boxes randomly, you will have to redesign and retest for every performance level you want to develop for. That's what I've been saying, no game developer wants to make more than one, so they just use one thread because that's what will run everywhere. But with the GPU, you KNOW the player will have enough power. Like I said a sub $100 card already has almost 1 TFlop of processing power.

 

[SNiiPE_DoGG]

the people who are really interested in running physics at high res and high graphics settings have the i7 or other processor power to run on the CPU, dual cores are dead to the high end market. - and they had 3500 boxes going at the end of the video (no ragdoll at that point)

 

[Benetanegia]

DirectX 11 has no physics APIs, only the capability to run physics code on any DX11 GPU using the compute shaders. The question is: Who is going to author the physics API for DX11? Once you have an answer to that, I think you'll have an answer for the most dominent phyisics API 10 years from now.


What has worked in the past and remains very popular is needs-based physics. That is, you code it in house and make it as accurate as deemed necessary. An open, free-to-everyone, physics standard isn't very likely.

It's becoming popular to make an in-house game engine, because almost all the developers are becomeing bigger and are creating more than one house or are buying others, etc. It could happen that they eventually do the same with physics, but as of now they are primarily looking into third parties for the most part. According to this at least: http://www.bulletphysics.com/wordpress/ - Scroll down to Figure 10, most popular libraries.

 

[Benetanegia]

the people who are really interested in running physics at high res and high graphics settings have the i7 or other processor power to run on the CPU, dual cores are dead to the high end market. - and they had 3500 boxes going at the end of the video (no ragdoll at that point)

I don't have an i7 nor I'm thinking in having one and I'm very interested in physics at high res. On the other hand as soon as GT300 launches, I will buy that or a HD5xxx depending on the pricing, performance etc, unless I don't find any game compelling enough, which sadly is more than likely.

Yeah 3500 boxes is a lot but in the one I posted there are clearly much much more than that. And as I said that's an i7 and we don't know if it's overcloked. The average gamer, even if it looks for performance in games, has a dual core, until recently most people used to recommend dual cores to everyone that wanted to build a gaming PC. That people who ask won't change their CPU anytime soon, but they will probably change their card.

 

[SNiiPE_DoGG]

its a 965 @ 3.2 - I remember from the original article that was posted. I think that was more boxes that the physics moving trash video you posted.... how can you know without a count?

 

[Benetanegia]

how can you know without a count?

Density and size of the boxes relative to the radius of the tornado. It's a bold calculation.

The 250.000 is said by the guy at some point, in the one with the car trail.

 

[jessicafae]

OK I checked the whole thread looking for this but could not find it.

OpenCL and DX11 compute shaders are similar, but different to CUDA and ATi stream.

OpenCL is designed for a "heterogenous compute environment" meaning many CPU, GPU, SSE vectors, Cell processors, Larrabee vector units, AMD bulldozer vector units and so on. It does "runtime" just-in-time-compiling (JIT) to determine available resources (and some load balancing) and is designed to provide a single programmer API and hide all the threading/hardware details. This means that any middle-ware or game engine coded to the OpenCL API will take advantage of ANY available resource whether it is some free unused SSE bandwidth, available CPU cores or available GPU shaders. Since it does this at runtime, it means that there is the possibility that one frame might use SSE for for some physics calculations, but the next fame that physics code might run on some GPU shaders. It also means that any code written to OpenCL will adjust to the hardware. If one computer has a simple core2 duo but a quad GPU, it will do most calculations on the GPUs. If a different computer has 24x i7 cores and a simple GPU more of the calculations will be done on the i7 CPU/SSE units, if the code is run on a PS3 the OpenCL will use the Cell processors. There is a developer at our lab porting his parallel/threaded/SSE code to OpenCL on OSX snow leopard right now and loves the API and the performance.

I think DX11 ComputeShaders is targeting the same "hetrogenous compute" API concept as OpenCL. It is possible to code an engine using DX11 graphics API and OpenCL compute API and skip the DX11 compute shaders.

CUDA and ATi Stream are GPU only APIs. Nvidia's recent OpenCL toolkit is GPU only. ATi's windows/linux OpenCL toolkit is currently only CPU/SSE but the CPU/SSE/GPU version will be out soon. Apple OSX 10.6 OpenCL toolkit is CPU/SSE/GPU right now and supports both Nvidia and ATi GPUs.

Havok has already ported sections of its toolkit to OpenCL.
ATi helped port the opensource Bullet physics toolkit to OpenCL.

And Nvidia is considering porting Physx to OpenCL
http://www.bit-tech.net/news/hardware/2009/03/27/nvidia-considers-porting-physx-to-opencl/1

Yes today Nvidia's Physx driver uses CUDA (or a CPU emulation layer in the Physx driver) to do Physx. But it is possible this may change. If Nvidia does not port Physx to OpenCL or DX11 compute (and off of Cuda), they may loose any momentum Physx has gained.

Personnally I have Mirror's Edge, I love the game and the Physx effects, but I am definitely annoyed that I had to buy an Nvidia card to get the Physx effects to work properly. Unfortunately we may have another couple years before Nvidia-only Physx/CUDA games disappear. It will probably be 2+ years before games using OpenCL middleware will hit the market.

Sorry I had hoped this post would have been shorter.

 

[Benetanegia]

I didn't knew Bullet was being ported to OpenCL already. That's interesting and could be the answer that so many people here are waiting for. Although I have always known that PhysX was going to be ported, because they'll have no option. I still think that most developers will go with PhysX or Havok at first, because support is probably much better.

 

[jessicafae]

There is also the Pixelux physics middleware toolkit which is being used by LucasArts. Pixelux is also being ported to OpenCL

http://www.youtube.com/watch?v=VAnXolinMPE
http://www.youtube.com/watch?v=VAnXolinMPE&feature=related

 

[EastCoasthandle]

the Confirmation Bias is strong with this one....

I wish I could find the demo of i7 running physics on Havoc with thousands of entities.... google is not turning it up though - was in the news like ~4 months ago IIRC

Are you looking for this:
video
video
video The camera shakes "no" at 31 seconds. Listen to what is said.

 

[MrMilli]

Are you looking for this:
video
video
video The camera shakes "no" at 31 seconds. Listen to what is said.

Well that's what i'm expecting from a physics engine in 2009 - for it to be multi-threaded.
With Mark Randel being the man behind the Infernal Engine, i didn't expect anything less from VELOCITY Physics.
Ghostbusters: The video game has huge amount of physics running at +30fps just on the cpu. These days you can have a quad core for less than $100 ... why not use that cpu power?
From what i have read in the past, developers preferred Havok over Physx. It seems to me that the two main reasons why Physx gained traction are: It's free and it's built into the UT3 engine.
I know i'm repeating myself but like i have said already, nVidia can multi-thread the cpu version of Physx like in no time. The PPU and GPU version are already multi-threaded.
The fact that nVidia acquired Ageia in February 2008 and had a demo already running in april 2008 on a GPU says it all.

 

[Benetanegia]

Well that's what i'm expecting from a physics engine in 2009 - for it to be multi-threaded.
With Mark Randel being the man behind the Infernal Engine, i didn't expect anything less from VELOCITY Physics.
Ghostbusters: The video game has huge amount of physics running at +30fps just on the cpu. These days you can have a quad core for less than $100 ... why not use that cpu power?
From what i have read in the past, developers preferred Havok over Physx. It seems to me that the two main reasons why Physx gained traction are: It's free and it's built into the UT3 engine.
I know i'm repeating myself but like i have said already, nVidia can multi-thread the cpu version of Physx like in no time. The PPU and GPU version are already multi-threaded.
The fact that nVidia acquired Ageia in February 2008 and had a demo already running in april 2008 on a GPU says it all.

As I see it, it is multi-threaded too. Some games do use more than one thread, like mirror's edge. I have already said this, but IMHO the limitation is probably on the game engine and not in the PhysX engine.

 

[RejZoR]

Funny. I thought Unreal Engine 3 is multi-thraded. And since Mirror's Edge is based on it, i find that kinda ironic.

 

[MrMilli]

Well RejZoR, i hope this answers your question:

http://www.anandtech.com/video/showdoc.aspx?i=3171&p=3

From source:

We also have a quick look at CPU usage for UT3 when running this map, to get an idea of just how well the CPU is being used; if it’s not being well used by the software PhysX simulations, then CTF-Lighthouse in particular is not a fair comparison. UT3 is designed around dual-core processors with some light helper threads to occupy cores 3 and 4, so if physics are the bottleneck then the multithreaded software PhysX simulations should be able to completely load the CPU.

However we are not seeing this. CPU usage is hovering at a little below 60% total usage on our QX6850. Given that we're not GPU limited, one thread in particular must be the CPU limited thread, and isn't capable of being further split. Our first thought is that the thread(s) that handle game logic also are doing the software physics, which would explain why we're not seeing a greater use of the other cores for the physics calculations. It's unfortunate if this is the case though, as it means that Unreal Tournament 3 (and possibly other Unreal Engine 3 games) is unable to fully utilize more than 2 cores and change. It also means that we potentially could be getting better software performance, which works against the PhysX hardware.

 

[FordGT90Concept]

It's becoming popular to make an in-house game engine, because almost all the developers are becomeing bigger and are creating more than one house or are buying others, etc. It could happen that they eventually do the same with physics, but as of now they are primarily looking into third parties for the most part. According to this at least: http://www.bulletphysics.com/wordpress/ - Scroll down to Figure 10, most popular libraries.

It only adds up to 63.9%. Where's the majority? In house/custom?

 

[Benetanegia]

It only adds up to 63.9%. Where's the majority? In house/custom?

First of all the rest is not the majority, ~64% is the majority. Nevermind.

The rest is a mix between many other engines, some in house some not, according to the site where I first saw that chart posted. They even named a few but I don't remember. But yeah the others where probably engines that had been used in less than 5 games, but third party still. Others are in-house and probably meant to be used in many games releasing in a short time period (so that technical improvement is not expected). It's not cost effective to design a new one for every game.

 

[FordGT90Concept]

I do believe a lot of that 36.1% is developers that don't use a third-party physics API. Instead, they just use a collection of static formulae integrated into their game engine at some point. They might use the same engine in multiple games but it doesn't fall under a separate physics engine or API.

 

[Benetanegia]

I do believe a lot of that 36.1% is developers that don't use a third-party physics API. Instead, they just use a collection of static formulae integrated into their game engine at some point. They might use the same engine in multiple games but it doesn't fall under a separate physics engine or API.

I don't doubt that many of that 36% is using propietary engine, but as you very well pointed out is probably more like some extensions added within their engine, which is used (or is going to be used) in all of their games. What I mean is that most of the games that use a propietary engine come from big comanies, that will use that engine in most of their games. Take Ubisoft for example, they release a lot of internal games and most of them are based on the same engine. Crytek also has 3 studios and a lot of working projects (as many as 4 apparently), which use CryEngine so it does make sense to make their own thing. But smaller studios* have to concentrate on the game and not in the tech.

Anyway, IMO a lot of that 36% also comes from games that don't require a complete physics engine, because of how the game is. Many strategy games for example.

*Doesn't mean their games are small. Think Bioshock, think Mass Effect...