Intel man touts real-time ray-tracing

Sounds like you'd need the rumoured Larrabee to do that ray-tracing, eh?

http://forum.beyond3d.com/showpost.php?p=1057810&postcount=21

interesting view from someone who is familiar with raytracing on gpu's...

If the rumours are true and Intel's graphics chips are based on x86 wouldn't the massive number of them (16+?) satisfy the requirements for realtime raytracing? Especially if it was disigned with this in mind. This was overlooked by Anandtech. You would not just use the CPU but you would use the GPU for most if not all the graphics processing.

More to the point, what about DAAMIT?

They are in the unique position of having both the x86 and the GPU under one roof. Im not exactly sure how that helps, but still they have to be in a better position than nVidia and Intel.

I've been raytracing in real-time for over thirty years now, and I only have one core! Booyasuks Intel! 8¬P

Thats incredibly selective imaging on the blog, you can recreate the reflection with a cubic-environment maps and caustics on the spoon would be far from useable in realtime-raytracing. Only thing i can think of that would be superior in realtime-raytracing is that you could have lovely perfectly smooth spheres with no poly-edges.

mmm, something is not plotting with this blog.

Have a look at the comparison picture. 

Seriously, a laymen might be impressed, but a gamer knows better. The effect on the kettle is called enviromapping. Geforce 2 could pull that convincingly back in 1999. It was one of the "features" back in the day.

Quake 4 was great but it is based on idTech 4, 3 years old. Lets face it, games have moved on. Crysis pushes probably 10x the detail that idTech 4 could handle.

By the time 2009 comes around with 8 core cpu's, game engines would probably again pushed 10x the amount of detail of Crysis.

I just don't see Intel breaking this cycle with a chip that have enough cores. 

Unless they build a dedicated ray-tracer. But if they are doing that, what is stopping nVidia or ATI to do the same?

Makes one wonder why Intel is building a GPU if they are considering then a dying tech?

Maybe The Inq should hold back from posting articles about this topic by authors who clearly leck the necessary knowledge.

"Since ray-tracing games allows for better/faster/more realistic graphics scenes, this means that Intel is on course for world domination.

The argument is compelling."

Not really. Just have a look around what is used nowadays in CGI production. Raytracing was the real deal back in the 90s maybe but since then we have much better methods to create realistic scenes. Like Radiosity or Image based Lightning. Raytracing is selectively used for those parts of the scene where there is no better way to get the desired results. For example for some refractive surfaces. But most of those are simulated anyway. BTW - you know all those great looking Pixar movies? None of them really uses raytracing. The first one where they implemented it in the in-house Renderman software is Cars.

About that Q4 raytracing demo. Did you actually look at it? The game looks better with the normal ID renderer then with the raytracing engine.

Onec more - please do some research before you fall for the propaganda. Maybe start by reading those comments for the linked Intel blog entry...

I love real time ray tracing, and have been following it in the demoscene for years... just check out the work of Realstorm or the amazing Heaven7 demo... (64k and 7 years old! Rasterize THAT)

@ Thinius - well done for using poor maths, flawed logic and arbitrary numbers to argue a point...

Environment mapping is a 'faked' effect - it impressed me intially, but on CS: Source for example (on gun scopes), with ray-tracing the effect isnt faked, so looks MUCH better. And works on actually curved surfaces.

You argue that Crysis is arbitrarily 10x as detailed as an engine 4 years old, THEN argue that it will only take another 2 years to become 10x again as detailed... Can you say random figures?

Ray-tracing is not a 'dying tech'. It is the technology used by Hollywood to add effects to their films, and even make films (Final Fantasy, Toy Story, Incredibles). This is because the image quality is better - IF you have the processing power...

Why does Intel always seem to forget to mention that GPUs are also increasing in speed at the same time as CPUs?


Also with regards to the comment on Beyond3D, AndyTX was actually generous towards raytracing. The O(n) sort does not guarantee a O(log n) search, though for most scenes it will be very close. To have the guaranteed O(log n) scaling would require a O(nlog n) sort.

Long story short raytracing is slower than rasterization, and it always will be.

can anyone remember how superior technologies have been scrapped because of cost. It does not make sence to make a game for 8 cores, ray tracing or any kind of exotic hardware. It might be nice for the media, since it gives them something to write. You automatically rule out a big source of revenue. If people are complaining about vistas requirement what makes you think they wanna ray trace? Nice, but how many of you have a gt40 or a mclaren f1?

Kind of funny reading how "a layman wouldn't know" but "a gamer knows better".

I don't think you DO know better. I'm also a gamer --- for about 30 years. I'm also a programmer. I'm very impressed by the quality of the ray-traced image - did you notice the self-shadowing on the objects? 

Seriously, generating a scene of comparable quality with directx is very cumbersome and difficult. I don't think I could do it. Do you have any idea how hard it is to generate volume shadows, self-shadowing, bloom, etc etc along with the myriad of other things that have to be done? Not to mention sorting polygons or even having multiple sort lists if you have transparent textures??? All that guff isn't necessary with ray-tracing.

I agree with the writer, this IS the future.

I want a real time raytrace/radiocity combination, but really, the effects done with rasterization are getting to be quite amazing.

Crysis being 10x as complex to render as Q4 sounds reasonable, but lets just go with 4x. It will be released about the same time as an 8 core gaming platform (yes 16 core intel systems are out, but those require Linux or Windows Server). Also, people buying such an expensive platform will want a screen that will do at least 2560x1600 and have a minimum frame rate near 60, which is probably about where the raytraced Q4 was. Intel will probably need about 16x the horsepower of their new, high-end gaming platform to raytrace a new, high-end game.

According to Moore's Law, Intel will be able to raytrace something as complex as Crysis in around 6 years (give or take a couple), in which time GPU's will be about 32x as powerful with even more rendering tricks and there will be another release or two of Windows non-Server. I don't think traditional x86 hardware can ever keep up. However, using a GPU to raytrace seems reasonable. I think the rumor was that the G92 chip had 1 TFLOP of processing power, and you can currently (or soon) buy a 1U 4 TFLOP NVIDIA supercomputer. Compare that to Intel's raytracing on around 55 GFLOPS.

Well its not just the hardware thats advancing, but also the techniques behind raytracing are advanicng at a similar rate. In a few short months intel themsevles imporved the framerate quite a bit and upped the resolution as well in this time (just check the difference between their last two demos of this). To those who think it doesnt look as good it is due to a variety of factors. Most notably is the fact it is a proof of concept demo and not a consumer product that needs to have every texture refined to the max.

I'd LOVE to have 16 cores - but with at least 4GB/s per core, otherwise we'll have bottlenecks.
And I don't like bottlenecks (unless I'm pouring myself another Scotch, of course).

There is no doubt that real-time raytracing is interesting. I do wonder though, why many of the unsolved problems of raytracing are not mentioned. Real-time raytracing can only work well if rays are cast in coherent bundles that also access memory in a coherent way. This is usually not true for rays refracted or reflected by e.g. curved objects. If ray coherence breaks down you are essentially limited by memory access speed. 

Also raytracing massively dynamic scenes is problematic. Usually dynamic objects are ratraced via fuzzy k-d-trees that need to be dynamically updated on the CPU once an object e.g. changes it pose. Imagine the memory needed and the processing time for updating the k-d- tress for all 10000 characters of a RTS game. Even given that all that scales nicely to a number of cores this is going to hurt performance a lot.

Another way to handle dynamic objects is to 'rasterize' their polygons into 3d grids (aka volume texture :-). These grids can then by raytraced. Again, the questions is how can this scale to a big number of uniquely animated objects.
Since 'rasterization' is something GPUs are good at it may turn out that upcoming GPUs can accelerate raytracing very well.

The intel blog seems to suggest that there is either raytracing or rasterization but nothing in-between. The thruth is that many GPU techniques already use raytracing or raycasting in shaders. So even at this point in time hybrid solutions are used. I would assume the future will show more of these hybrid solutions and that the future itself is a hybrid solution.

To conclude I would like to mention, that games have always been about 'good enough' graphics. There is no need for an exact image without errors if the errors are small enough. In most cases the correctness of raytracing is not needed at all.

I guess, GPU vendors won't let steal their food. If processing power is good enough for ray tracing, the GPUs will introduce ray tracing enhancements. That might convert raster cards to ray tracing cards. As CPU cores have to go a long way for dozens of cores, GPUs l evolve and support hardware ray tracing, radiocity, etc. The market will shade the GPU functionality - it won't be revolutionized as there are to many players with different interests. And that's good.

I looked up this stuff a few weeks ago. This "State of the Art" publication resolved most of my questions:
http://www-csl.csres.utexas.edu/gps/publications/eg07-STAR-dynrt/

To summarize, much of the current research in RTRT has to do with finding more optimal algorithms for culling; the layperson RT implementor mostly knows of k-trees(a subset of BSP) as "the solution," and this is so because tree-building time was not as much of an issue when the available power only enabled off-line rendering. Variations on k-trees have come up in the research, but a number of alternative methods have also been attempted with some success. The paper describes some examples of highly dynamic animated scenes using bounding volume hierarchies, for example.

Long story short, I wouldn't believe the detractors yet. Rasterization is still ahead for the time being, but it's entirely possible that the RTRT researchers will make an algorithmic breakthrough in the coming years.