Could PowerVR be adding real-time hardware ray tracing to their 3dchips?

[Baboonanza]

Well, breakthroughs in algorithmic efficiency (where a floating point utilisation saving can be made) are always more valuable than "brute force" hardware improvements. That being said, in my experience massive algorithmic performance improvements tend to come with quite severe restrictions on their scope of applications, in comparison to the traditional algorithm they are replacing (I'm talking in general terms about numerical methods here, not specifically about ray tracing).

Still, it's an interesting development that is well worth watching. But algorithmic improvements or not, ray tracing is still heavily reliant on floating point power, and massive floating point performance is extremely difficult to scale down to low-power devices. For that reason, I can't see it finding a home in mobile devices before it is implemented on desktops.

And to make ray-tracing practical on a mobile device would require a performance improvement of literally 100s if not 1000s of times. And since ray-tracing isn't one single algortihm it would require multiple, significant breakthoughs. I just don't see it. More likely to me is that some part of this process is relevant to PowerVRs rendering engine, perhaps the optimised ray-casting altgorithm that can be used to speed up the tile-based rasterizer, and they wanted that piece specifically. Or they want the IP for non-mobile product development.

Also, while digging for further information in their site I came across this from SIGGRAPH 2009:

The scene shows the highly flexible OpenRL API in action with a multi-million polygon architectural scene. Over 80 complex materials, all created in Autodesk 3Ds Max Design and automatically converted to standard OpenGL Shading Language (GLSL) are required to render the scene. Many highly incoherent secondary rays are cast through the scene demonstrating effects such as shadows, high dynamic range lighting, spherical lens camera and glossy reflections all in real-time. Scene dynamism is demonstrated showing geometry, shader, lighting and camera changes all occurring without work-flow interruption. At VGA resolution the scene renders at approximately 5 frames-per-second, with raytraced HDR, Domelight, Reflections, Refractions, etc.

So: Their own demo,full-sized custom add-on board, VGA resolution (800X600 max) at 5 times frames a second.

And you think they are going to bring ray-tracing to mobile devices? Dream on.

 

[Pottsey]

Baboonanza said “Their own demo,full-sized custom add-on board, VGA resolution (800X600 max) at 5 times frames a second.
And you think they are going to bring ray-tracing to mobile devices? Dream on.”
The first gen dev board from 2009 was 5 FPS the next gen broad from 1010 was x14 faster or so they say which is around 50 to 70 FPS. Assuming they keep the current pace and have one generation a year even if it’s less than a x14 improvement then by 2011 we should get a 3rd gen board with just enough speed needed for games. But we won’t be seeing a PowerVR chip with this till a 4th gen in 2012 at the very easiest. It’s not hard to imagine by 2012 or 2013 a PowerVR chip with the power to pull this off. It’s not as far out as you might think. I would be very surprised if this arrives after 2015. My guess is between 2012 and 2015 with it most likely being 2013 or 2014.

The IP is very clearly for both mobile and none mobile chips. The CEO said they got to tech for handhelds and consoles.




Baboonanza said “And to make ray-tracing practical on a mobile device would require a performance improvement of literally 100s if not 1000s of times. And since ray-tracing isn't one single algortihm it would require multiple, significant breakthoughs.”
Not sure where you get those numbers from. They only need about x20 performance over the 1st gen chip and only about x3 performance over the 2nd gen chip. Not 100’s or 1000’s of times more performance. A x3 performance over 2 to 3 years should be doable.

I don’t deny that this is still years away, but I believe it’s much closer than most people think.

 

[Duff-Man]

They only need about x20 performance over the 1st gen chip and only about x3 performance over the 2nd gen chip. Not 100’s or 1000’s of times more performance. A x3 performance over 2 to 3 years should be doable.

The first-gen hardware can do ~5fps on a VGA (600*400) resolution image. To get ~50fps on a 1080p (1920*1080) image, assuming perfect linear scaling, you're already looking at a factor of 80+ improvement.

Secondly, consider that the number of polygons used in game scenes is likely to be significantly higher than in the sample images (particularly by the time ray tracing becomes available). For example, images like this and this that are rendered at around 5fps on the first-gen hardware are impressive, but geometrically they are not particularly complex. Increasingly complex geometry requires increasingly more computing power with ray-tracing (sure it scales linearly rather than logarithmically, as with resterisation, but the number of rays will be far greater than the number of raster light sources).

Anyway, we're not a million miles off, but it's certainly not just "another factor of three" on the current-gen hardware, even assuming that the "factor of 14 speedup" is truly representative (as opposed to being 'in a best case scenario' - we really don't have enough detail to say for sure).

I don’t deny that this is still years away, but I believe it’s much closer than most people think.

I don't want to seem like I'm pouring cold water on the idea - I'd love to see the capability for ray tracing sooner rather than later... But experience has taught me that when it comes to technological development, a slightly pessimistic viewpoint tends to relate more closely to reality than a slightly optimistic one. People will always present the best results that they can (I'm as guilty of this as anyone when it comes to my own work), and gloss over the areas where their solution currently falls down, or needs more work.

 

[Pottsey]

Sorry my post wasn't clear. When I said a factor of 20 I meant for a 1024x768 res for mobile screens.

As for more complex screens needing more power I was wondering about how that would work with tile based cards that remove all hidden surfaces. I assume the rays would have to be in the pipeline before the objects are removed and cut out so they are not rendered.

I was think thinking you could have fun stuff like a ray traced photo realistic head with a real photo textured on of freinds for talking head when you ring.



MoodyB said "http://www.pixelux.com/dmmEngine.html
How would you class DMM then ?"
Not sure I follow; DMM is not hardware physics so it doesn't fit into any of the options I posted. The options I posted are for hardware physics only.

 

[Duff-Man]

Sorry my post wasn't clear. When I said a factor of 20 I meant for a 1024x768 res for mobile screens.

Fair enough, but I still maintain that improving the performance of a desktop-scale product by enough to work with desktop resolutions is going to be A LOT less challenging than scaling down the size and power-consumption of the processor to the standard required by mobile products. For battery-powered mobile products, you really need to be looking at <1W power consumption. For fast parallel computing devices this is extremely difficult to achieve...

Okay, as a slightly more tangible example, consider 3D gaming: At a lower resolution, you use a lot less rendering power... But, games on low-res mobile screens are nowhere near as sophisticated as those on desktop machines, despite the resolution being (say) a factor of eight lower. Why? Well, because (just like ray tracing) 3D graphics rely on fast parallel computations, and it's extremely difficult to get fast parallel computations with such a low power draw. If this was not the case, then we would be already be playing games with (say) 2005-era graphics on our mobiles.


Put another way: If we can't do on mobiles today what we could do 5 years ago on a desktop (as far as graphics are concerned), then there is no way it's going to be easier to do ray-tracing on a mobile than it is on a desktop.

As for more complex screens needing more power I was wondering about how that would work with tile based cards that remove all hidden surfaces. I assume the rays would have to be in the pipeline before the objects are removed and cut out so they are not rendered.

An interesting point, and frankly I have no idea

Certainly some culling would have to take place, but exactly how this would be achieved while saving on rendering un-needed rays... Well, it would be more complex than simply culling textures or polygons, because rays "bounce around", and a 'hidden' polygon may still influence a ray in some way. In a way, you need to render the ray in order to know whether it should be culled or not!

I'll leave figuring this out to the poor guys who have to work on this nasty optimisation problem

 

[MoodyB]

Not sure I follow; DMM is not hardware physics so it doesn't fit into any of the options I posted. The options I posted are for hardware physics only.

It was more from the point that with tools like DMM, Euphoria, Bullet, Havok & ODE there's no real need for hardware based phsyics, in games at least.

And what do you class as 'hardware physics' exactly ?

I take it you mean hardware accelerated, as in PhysX with the ageia ppu / Nvidia gpu ?

 

[Pottsey]

MoodyB said " It was more from the point that with tools like DMM, Euphoria, Bullet, Havok & ODE there's no real need for hardware based phsyics, in games at least."
I disagree as all those are limiting us big time. Many physics effects in games are skipped or scaled down looking worse because we are limited by software solutions. The CPU just cannot handle high end physics forcing many games to cut corners or scale effects down.

If everyone had hardware physics we would have far better physics then we have today. As for what I consider hardware physics, well anything done in hardware like a PPU or GPU

EDIt: Take this from the 2011 Eve online expansion http://www.youtube.com/watch?v=mrtwESnTOwY&feature=player_embedded
Hair and cloth/clothes like that on a CPU just doesn't work at useable speeds. Now add in that she's going to be in a star base in an MMORPG with lots of people like that and you need hardware physics. Then you have liquids which just don't work well on the CPU in real-time.

 

[MoodyB]

If everyone had hardware physics we would have far better physics then we have today

Well imho until the gfx card vendors put away the handbags & agree on a single standard, or M$ incorporate a physics engine into DirectX, that isn't going to happen.

Maybe the best chance is when Bullet / Havok have OpenCL support, so everyone can benefit ?

Edit: Take this from the 2011 Eve online expansion http://www.youtube.com/watch?v=mrtwESnTOwY&feature=player_embedded
Hair and cloth/clothes like that on a CPU just doesn't work at useable speeds. Now add in that she's going to be in a star base in an MMORPG with lots of people like that and you need hardware physics. Then you have liquids which just don't work well on the CPU in real-time.

Do we know that uses the gpu, as Apex is cross-platform & will have to use the a single core / thread on consoles I'd imagine

The Havok cloth demos afaik run on the cpu & don't look that different to the above video. I could be wrong though.

 

[Rroff]

Do you have any info on this? It doesn't seem realistic that a device from back then could perform real-time ray tracing... At least not on a large scale.

That said, if it was a proof of concept device back in 1999, it could well have just been a demonstration of a flexible parallel compute device (similar to the programmable GPUs of today), which implemented a cut-down ray-tracing algorithm to demonstrate the potential of such technology (...and the improvement relative to CPU rendering).

Its a long time ago, think it was mentioned in PC Zone mag, it could do 50 million linepicks a second with 10K triangles if memory serves, pretty sure it didn't do caustics or anything more complex tho.

EDIT: Might be getting the stats mixed up with one of the voodoo card specs lol but that sounds about right.