NVIDIA and Havok Demonstrate World's First GPU-Powered Game Physics Solution at Game Developer's Conference

[DaveBaumann]

Originally posted by: Gstanfor 1) pixel shaders already get loaded up pretty heavily already by developers and using them for physics as well would certainly lead to slowdowns in modern games making SLI almost essential (and even then there would probably still be a slowdown).

There are many rendering factors that will affect performance and generally speaking, the ALU processing capability isn?t really that stressed ? this is evidenced by the 3x shader difference between R580 and R520 but the actual performance increases in game being significantly lower (while it is evidenced in specific shader tests). There is more math power in the pixel shader pipeline, and its will continue to go that way (unless everyone unifies, in which case it becomes immaterial in future hardware).

2) Physics affects in game objects position and shape, things that get dealt with in the GPU by the vertex processors. It makes no sense to calculate this sort of info in pixel shaders then have to transfer it somehow to the vertex shaders so it can actually be used.

Actually what you?ll be doing is creating the calculations that will directly alter the positions of vertex data in vertex buffers, ergo its probable that you?ll want to be using a part of the pipeline that has direct write access to those vertex buffers in memory.

3) When you are dealing with physics in the gameworld you will want good precision in your maths otherwise you will end up with errors with object/vertex placement. Vertex shader math is more precise than pixel shader math because it has to be (uses FP32 exclusively).

All SM3.0 pipelines have to have FP32 processing capabilities. G70/G71 has nearly 5x the FP32 processing capacity in the pixel pipelines than it does in the vertex pipeline. You?ll be slowing things down far more if you use the VS.

 

[Gstanfor]

Originally posted by: DaveBaumann

Originally posted by: Gstanfor 1) pixel shaders already get loaded up pretty heavily already by developers and using them for physics as well would certainly lead to slowdowns in modern games making SLI almost essential (and even then there would probably still be a slowdown).

There are many rendering factors that will affect performance and generally speaking, the ALU processing capability isn?t really that stressed ? this is evidenced by the 3x shader difference between R580 and R520 but the actual performance increases in game being significantly lower (while it is evidenced in specific shader tests). There is more math power in the pixel shader pipeline, and its will continue to go that way (unless everyone unifies, in which case it becomes immaterial in future hardware).

2) Physics affects in game objects position and shape, things that get dealt with in the GPU by the vertex processors. It makes no sense to calculate this sort of info in pixel shaders then have to transfer it somehow to the vertex shaders so it can actually be used.

Actually what you?ll be doing is creating the calculations that will directly alter the positions of vertex data in vertex buffers, ergo its probable that you?ll want to be using a part of the pipeline that has direct write access to those vertex buffers in memory.

3) When you are dealing with physics in the gameworld you will want good precision in your maths otherwise you will end up with errors with object/vertex placement. Vertex shader math is more precise than pixel shader math because it has to be (uses FP32 exclusively).

All SM3.0 pipelines have to have FP32 processing capabilities. G70/G71 has nearly 5x the FP32 processing capacity in the pixel pipelines than it does in the vertex pipeline. You?ll be slowing things down far more if you use the VS.

Dave,

I never said competing vertex shaders don't use FP32 did I?

Also, about your SM3.0 comment up above, you do realise SM covers both pixel and vertex models don't you? SM3.0 doesn't automatically mean pixel shader.

I don't expect nVIDIA to enable physics support for nV3x, but, it certainly wouldn't be out of the question IMO given the capabilities of the nV3x vertex processors.

As I said in the quote above, I think the pixel shaders will have enough work to be getting on with, without being further burdened with extra load, and you still have the problem of how to get the results they compute to the vertex processors where it is useful.

 

[nts]

Originally posted by: Gstanfor
...and you still have the problem of how to get the results they compute to the vertex processors where it is useful.

You may not know this but video cards can read from and write to memory (RAM). They also come with some memory onboard which is much faster than using system memory (this is what your CPU uses).

So knowing that you have a place to store data (memory) and can read from and write to this magical place (memory again), do you see how you can solve the above problem or do you still need help?

 

[Gstanfor]

That would involve an unnecessary slowdown though nts and quite possibly, cpu involvement. Perhaps you missed this bit of the press release -

Utilizing Havok FX and NVIDIA graphics technology, game developers can now implement sophisticated physical phenomenon such as debris, smoke, and fluids that add immense detail and believability to game environments. Game designers can include advanced physics effects without burdening the CPU and slowing game-play, since the effects are simulated and rendered on the GPU.

Of course, it's entirely possible that if ATi does a Havok implimentation or something similar, it will use the pixel shaders rather than the vertex shaders since that would probably suit its architecture better. However in the end the data affects polygons and vertices not pixels.

 

[DaveBaumann]

I never said competing vertex shaders don't use FP32 did I?

? Nowhere was it indicated that you did say that.

Also, about your SM3.0 comment up above, you do realise SM covers both pixel and vertex models don't you? SM3.0 doesn't automatically mean pixel shader.

Yes?
The point being, though, that PS have to have full FP32 capabilities, so it has the same precision as current vertex pipelines, and you also have a lot more of it.

As I said in the quote above, I think the pixel shaders will have enough work to be getting on with, without being further burdened with extra load, and you still have the problem of how to get the results they compute to the vertex processors where it is useful.

Vertex Shaders read from vertex buffers ? the point being is that your vertex data will reside in the vertex buffer, there will be a ?physics pass? that can calculate and manipulate that buffer, write it back out so that the "graphics pass(es)" will then process the data from the manipulated vertex buffer.

 

[nts]

Originally posted by: Gstanfor
That would involve an unnecessary slowdown though nts and quite possibly, cpu involvement.

WTH are you talking about?

How much simpler can I make it for you.

1. Pixel shader writes to memory
2. vertex shader reads from memory
3. Done, NO CPU Involved.

Graphics card memory also stores geometry information and a lot of other info besides textures.

 

[Gstanfor]

YOur FP32 point was in there before you conveniently editted it out Dave...

 

[RobertR1]

Originally posted by: Gstanfor
YOur FP32 point was in there before you conveniently editted it out Dave...

Sucks to be slow at quoting

 

[DaveBaumann]

The only thing I edited, Greg, was "...was direct write access to those vertex..." - I added 'write' there to clarify.

 

[Gstanfor]

My gut feeling is still that you can't put too much physics workload on the pixel shaders, they already have enough to deal with. Making a game perform slower isn't going to help adoption and if you require SLI for it to be useful you may as well buy a dedicated physics card and be done with it - you won't have to pay for expensive RAM and unnecessary features that way. Bear in mind this physics feature isn't just for GF7 either, they are making it available for GF6 as well.

 

[RobertR1]

Originally posted by: Gstanfor
My gut feeling is still that you can't put too much physics workload on the pixel shaders, they already have enough to deal with. Making a game perform slower isn't going to help adoption and if you require SLI for it to be useful you may as well buy a dedicated physics card and be done with it - you won't have to pay for expensive RAM and unnecessary features that way. Bear in mind this physics feature isn't just for GF7 either, they are making it available for GF6 as well.

wait, wait, wait...can we get back to Dave's quote editing thing? So did he or did he not edit out the FP32 part???

 

[akugami]

Originally posted by: Crusader

Originally posted by: akugami
http://forums.anandtech.com/messageview...&STARTPAGE=4&FTVAR_FORUMVIEWTMP=Linear

Crusader, read that thread and read my post in there. I did a little searching and it turns out ATI sells nearly as much discrete GPU's as nVidia meaning that the popularity level of ATI's Radeon video cards are about on par with nVidia's Geforce video cards. Basically it shows that even with the superior performance of nVidia's Geforce 7 cards over what ATI had out at the time, ATI was still doing quite well. If you got numbers showing differently then please do so by all means.

Dead wrong.
I was talking about current gen video card sales. Not integrated, not yesteryears. This is X1900 vs GF7 that I was referring too. Sorry. It has nothing to do with what I said.. therefore you are posting ATI fanboy garbage if you are directing your info towards me.

Originally posted by: Crusader
No I'd say the GF7 series has massively more popularity than the X1900s.

Try again next time.
You knew better than to try to post that crap as a rebuttal to my statement. See ya.

You say I'm wrong, show me reports refuting my thoughts. I'm dead wrong, but you have no proof I'm wrong? Again, the only report of nVidia dominating ATI to any large degree was last seen in late 2003 by my searches.

In 2004 and 2005 ATI took roughly half of the GPU market and in certain segments of the market such as integrated notebook GPU's, it took 60% and recently surged to 77%. In discrete desktop sales it alternated between 40 odd % to 50 odd %, which is the same fluctuation my searching showed nVidia with. Meaning no one company dominated, contrary to your and many other people's claims.

Now, there aren't a lot of numbers for 2006 yet since Q1'06 is not even over yet but there is nothing to suggest that the numbers would be remarkably different. I'd wager that if in 2005 with nVidia having nearly an 8 month undisputed period where it was top dog and the market was still split roughly even between nVidia and ATI what makes you think ATI will sell drastically less video cards than nVidia when the overall performance between the top two GPU's of each company is nearly evenly split?

I'm not saying it's outperforming nVidia but it's more than holding it's own and at worse, in the last two years it's selling 40% of all video cards sold. That is insanely good when you are doing 40% of the business in a market. Ask Samsung, Cowon, iRiver, Creative, etc, if they were to have 40% of the DAP business by years end with Apple owning 55% if they would be happy with this or not? I'd wager any amount you care to place on the table that they'd be ecstatic. Forget 40%, ask any company if they were to have only 30% of a multi billion dollar market if they'd be happy or not.

It all comes down to this, I have links to reports showing my claim that ATI is doing very well. From all indications it's holding roughly half the market between it and nVidia. That's not counting total GPU sales which it holds 26% compared to nVidia's 18% with Intel the overwhelming market leader. All I'm saying is that ATI is not getting owned and holding it's own. I'm not going to make outrageous claims about it kicking nVidia's rear end but from what I can see and what I've dug up I don't see it being slaughtered by nVidia. Is it so hard to accept that ATI as a company is doing quite well? Is it so hard to accept that their market share will not suddenly decline overnight when the products they market are now much more competitive compared to their competition's products in the last 8 months? It's simple economics, if economics can be said to be simple.

Again, you say I'm wrong but you have zero proof to back up your words. Prove me wrong. Show me reports where nVidia's sales numbers are overwhelming ATI's.

 

[DaveBaumann]

As I said, its likely that this is going to operate as a rendering pass within a frame ? i.e. when a frame is rendered its not often that it just renders straight to a colour buffer and then displayed all in one. Take Doom 3, for instance, it has a Z buffer pass, that renders all the geometry just to lay down Z information, then it has multiple geometry passes for each of the light to calculate the shadow buffers, then it might get round to a standard colour pass, before blending that colour buffer with the shadow buffer data.

What we are talking about here is adding an extra pass in there, so along with these types of rendering passes there will also be a ?physics calculation pass? that calculates the collisions for that frame, alters the vertex buffer(s) accordingly so that the subsequent passes can read from that buffer and calculate the geometry based on the ?post collision detection? vertex information.

For this very reason, its probably that its not going use a single element of the pipeline as that would actually slow down the physics rendering pass ? the pre collision geometry will need to be calculated in the vertex shaders (as normal) and then that rendered through the pixel shaders with the collision program and the calculated collision data rendered to a vertex buffer.

 

[nts]

Originally posted by: DaveBaumann
...
What we are talking about here is adding an extra pass in there, so along with these types of rendering passes there will also be a ?physics calculation pass? that calculates the collisions for that frame, alters the vertex buffer(s) accordingly so that the subsequent passes can read from that buffer and calculate the geometry based on the ?post collision detection? vertex information.
...

You don't need full vertex info either, think bounding volumes, represent them with equations even and just update the final transform matrix

 

[akugami]

Originally posted by: Gstanfor
My gut feeling is still that you can't put too much physics workload on the pixel shaders, they already have enough to deal with. Making a game perform slower isn't going to help adoption and if you require SLI for it to be useful you may as well buy a dedicated physics card and be done with it - you won't have to pay for expensive RAM and unnecessary features that way. Bear in mind this physics feature isn't just for GF7 either, they are making it available for GF6 as well.

I'd have to agree with this but I think it applies to both ATI and nVidia. I don't think without a core designed to handle physics the GPU's are not going to do physics too well. If it's the pixel shaders that are doing most of the work however, it would stand to reason ATI would be better at this than nVidia.

It'll be interesting to see how the G80 core will be. Will it have it have a physics coprocessor or more shader pipelines? I'm wondering if future GPU's from both companies will be multicore chips. Basically the main GPU along with a specialized physics unit on the same core. This would make the most sense than just throwing more shader pipelines out there.

 

[Gstanfor]

Yes, I fully understand that Dave, but looking at how current and previous games behave when you up the pixel shader load on them I find it quite difficult to believe there won't be a negative impact on framerate. At the end of the day, if you give the pixel processors more work to do, they are going to require more time to accomplish their tasks, no matter how many passes you create. Computing time isn't free.

If GPU based physics acceleration requires SLi or crossfire to be genuinely useful then it won't be of much interest to me.

Remember the S3 Virge - no-one appreciates a decellerator no matter how cool its functions may be

 

[DaveBaumann]

Does having the shadowing effects in Doom 3 make it slower than not having them? Does having FSAA enabled make it slower than not? Does having high quality AF filtering make it slower than than bilinear filtering? Does having high quality SM3.0 effects make it slower than having simple bumpmapping and multitexture effects? Simply put, when has affecting effects not had a negative effect on a piece of hardware and when have the vendors stopped releasing higher performance hardware.

At the end of the day, graphics vendors want to sell graphics cards - Physcis rendering on the graphics processor is just another process that will increase the quality of the rendered output, much like many of the other features they provide, and another way of selling you a graphics card. They would sure as hell like you to buy the latest chips they have produced rather than adding a separate piece of silicon for physics processing.

 

[DaveBaumann]

Originally posted by: nts
You don't need full vertex info either, think bounding volumes, represent them with equations even and just update the final transform matrix

Yep, indeed.

 

[Gstanfor]

Originally posted by: DaveBaumann
Does having the shadowing effects in Doom 3 make it slower than not having them? Does having FSAA enabled make it slower than not? Does having high quality AF filtering make it slower than than bilinear filtering? Does having high quality SM3.0 effects make it slower than having simple bumpmapping and multitexture effects? Simply put, when has affecting effects not had a negative effect on a piece of hardware and when have the vendors stopped releasing higher performance hardware.

At the end of the day, graphics vendors want to sell graphics cards - Physcis rendering on the graphics processor is just another process that will increase the quality of the rendered output, much like many of the other features they provide, and another way of selling you a graphics card. They would sure as hell like you to buy the latest chips they have produced rather than adding a separate piece of silicon for physics processing.

You may want to check out a little utility for Doom3 called 3mood nvnews link download link.

It clearly demonstrates that as you up the shader workload the framerate will come down, regardless of shadow settings or anything else, and as I've told you in these forums previously it shows Doom3 can run in partial precision mode (FP16) with no IQ loss too...

EDIT: added download link for latest version of 3mood.

 

[Gstanfor]

Originally posted by: DaveBaumann

Originally posted by: nts
You don't need full vertex info either, think bounding volumes, represent them with equations even and just update the final transform matrix

Yep, indeed.

All depends on how accurate you want or need a particular physics scenario to be. There are cases when this is fine and cases where you want to work with the vertice or vertice mesh.

 

[imported_JamesDax]

I don't care what Nvidia or ATI does. Ageia's PhysX PPU is a better solution.

 

[dug777]

Originally posted by: Crusader
Shhh.. quit making sense GStanfor. This is not the place.

Shhh stop trolling, believe it or not, this is not the place...Gstanfor is quite capable of conducting a mature discussion (and he has in this thread) without you buzzing around sticking up for him, trolling & best of all, contributing nothing of any relevance to the thread topic :roll:

I suspect if we went back through your 216 posts we wouldn't find a single one saying ANYTHING complimentary about ATi, but then looking at your sig i suspect we're either either dealing with an AEG shill or the most dedicated Nvidia fanboy known to mankind

On topic, as i have been saying, regardless of whether it's pixel or vertex pipes, i can't see how this will be viable on exisiting hardware without a serious FPS drop...it's all very well saying that the GPU could use 'unused' portions of the core but
1) most people are GPU limited so at least one aspect of the pipes is going to be full at any one time...
2) GPU load from 'traditional' work is variable & thus this 'free' space is variable (if it even exists consistently at all), which would make the amount of physics work the GPU can do variable...
3)will CPU overhead in coordinating the GPU work rise? i assume that this would be less than the benefit provided otherwise it's a dead end...

Interesting stuff

 

[imported_Crusader]

Originally posted by: akugami

Originally posted by: Crusader

Originally posted by: akugami
http://forums.anandtech.com/messageview...&STARTPAGE=4&FTVAR_FORUMVIEWTMP=Linear

Crusader, read that thread and read my post in there. I did a little searching and it turns out ATI sells nearly as much discrete GPU's as nVidia meaning that the popularity level of ATI's Radeon video cards are about on par with nVidia's Geforce video cards. Basically it shows that even with the superior performance of nVidia's Geforce 7 cards over what ATI had out at the time, ATI was still doing quite well. If you got numbers showing differently then please do so by all means.

Dead wrong.
I was talking about current gen video card sales. Not integrated, not yesteryears. This is X1900 vs GF7 that I was referring too. Sorry. It has nothing to do with what I said.. therefore you are posting ATI fanboy garbage if you are directing your info towards me.

Originally posted by: Crusader
No I'd say the GF7 series has massively more popularity than the X1900s.

Try again next time.
You knew better than to try to post that crap as a rebuttal to my statement. See ya.

You say I'm wrong, show me reports refuting my thoughts. I'm dead wrong, but you have no proof I'm wrong? Again, the only report of nVidia dominating ATI to any large degree was last seen in late 2003 by my searches.

You are dead wrong. I never said a thing about total sales. I said GF7 popularity vs X1900. I even quoted it for you in my last reply but you still are too dim to get it.
You are arguing something completely different than what I claimed in this thread.. and you so poorly missed the point too and now you are trying to argue something completely different.
Thats too bad.
You might be right in what you are saying in your points, I never said you werent.. I dont really care. But it has absolutely no relevance to my claim that you were attempting to rebut.

Or perhaps you were just trying to argue with me. You arent really arguing with me at all, because you answered the wrong point (one that wasnt even mentioned in this entire thread).
Reading. Comprehension.

 

[akugami]

Originally posted by: Crusader

You are dead wrong. I never said a thing about total sales. I said GF7 popularity vs X1900. I even quoted it for you in my last reply but you still are too dim to get it.
You are arguing something completely different than what I claimed in this thread.. and you so poorly missed the point too and now you are trying to argue something completely different.
Thats too bad.
You might be right in what you are saying in your points, I never said you werent.. I dont really care. But it has absolutely no relevance to my claim that you were attempting to rebut.

Or perhaps you were just trying to argue with me. You arent really arguing with me at all, because you answered the wrong point (one that wasnt even mentioned in this entire thread).
Reading. Comprehension.

Perhaps I'm too dim to get this. But let me see if I can get this straight. Sales of discrete nVidia GPU's and sales of discrete ATI GPU's are nearly equal. Ok. That means that not counting integrated GPU's, nVidia and ATI sells nearly equal GPU's. If we count integrated, ATI sells a LOT more GPU's. Ok, everyone following so far? That means that with some fluctuation, ATI and nVidia should sell nearly the same amount of video cards in each product range. That means that nVidia's Geforce 7 cards are more popular because they sell in roughly the same number as ATI's X1900's? I think you lost me there. Is this the Chewbacca Defense? Cause it doesn't make sense.

You have provided zero proof that starting from the time of the debut of the X1900 that nVidia sells more Geforce 7's than ATI does X1900's. If you can, I'd love to see where you get those numbers from. Because right now, I CAN prove that ATI and nVidia are roughly evenly split along sales numbers for discrete GPU's. This stands to reason that with some fluctuation aside all 3 ranges of cards should be roughly equal as well. I have given proof of what I say while yours is based on opinion as you have provided zero proof of anything you say. Is it conclusive proof? Absolutely not, for all I know you could be right in that on the high end, nVidia is selling more GPU's than ATI. But by the same token since we don't have any reports on a complete breakdown of low, mid, and high range video cards, only a total sales number and marketshare number, it could very well be that ATI is the one that dominates the high end of discrete GPU sales while nVidia dominates mid and low range sales. If you think I'm wrong, provide proof that I'm wrong.

Regardless, it is your opinion that nVidia is selling more Geforce 7's and it is your opinion that the Geforce 7 is more popular than the X1900. If it is fact that the Geforce 7 series is selling more than the X1900, please provide sales numbers and the web site you got them from. Or is it just another opinion you pulled out of the air and passing it off as fact? I mean, I ask for proof and you give me opinions and pass it off as proof. And you say I need to brush up on my reading comprehension? Maybe you need to look up opinion and fact in the dictionary when you head off to your reading comprehension classes. After the reading comprehension classes you might want to take up some logic classes and then reread what you wrote. Not responding to this issue anymore in this thread unless you can provide some solid facts.

 

[imported_Crusader]

Wrong again.

There is no "proof" yet on the outcome of the X1900 vs GF7 battle. It appears though the GFs are selling out, while the X1900s seem in stock.. the GF7 are rumored to be easier to produce.. yet selling out? People are clamoring about them on forums like this one?
Everyone wants one. Its another solid Geforce product with SLI support.

We wont know for sure the actual sales.. well.. you guessed it, till the next gen.. but its pretty clear that the GF7 is more popular than the x1900s and you have presented no data that destroys that claim.
I never presented data to back it up, because it cant be backed up yet.. the battle isnt over yet. Yet you bring in all these old statistics like you are proving something?
You arent! And I've been saying that to you repeatedly!