Sunday, December 6, 2009

Larraboom !!

Some pretty drastic news coming from Intel today.

It seems that Larrabee is being "refactored" so to speak. In practice, the excessive delays have probably rendered the hardware obsolete before even being released.
The BSN site has an interesting article with a timeline based on rumors and informants and even with a picture of what seems to be a Larrabee card.

Intel says that the LRB is going to be an SDK for the time being while the hardware is possibly being reworked... apparently even to the point of rendering the current SIMD instruction set obsolete ! much for my (fairly small) effort into writing the 16-way SIMD intrinsics version of the vector class template used in RibTools.
In any case, it was somewhat inspiring to see how a wider SIMD ISA looks like and how it should be used.

I don't want to be negative, but this "it's only being postponed, not canceled" talk is hard to believe. It reminds me of the Commodore Amiga fate.. where people kept hanging on promises for ages.

I'm not sure how it is with hardware.. but when it comes to drivers writing, the idea of having to support DirectX 9, 10, 11, OpenGL and possibly OpenCL.. and doing all this on a completely new architecture, while trying to beat nVidia and AMD/ATI cards (that have the luxury of only having to continue to evolve)...  that's a pretty huge task.
You can't really substitute people with a lot of experience in a specific subfield with a lot of generally smart people... (where by people I mean engineers 8)

I should really start thinking about looking into OpenCL/CUDA (and in that case, buy an nVidia card for my home PC).



  1. Yeah, unfortunately delays often lead to cancelation.

  2. Intersting how it almost exactly coincided with IBMs decision to discontinue the Cell.

  3. Davide, I need your help. It's not about Larrabee.
    I wish I knew if there is any possibilty to reference another mesh from a RM surface shader, in a single render pass. As I've found, a surface shader is attached to a mesh. I need within that shader to get the current shading point distance to the next mesh. Is that possible?
    Thank you very much!

  4. Hey Marius 8)

    Honestly, I only very little about RM shaders.. the little I learned so far was to get the basic compiler going (^^;)

    I think that there is a trace() function, but apparently PRMan doesn't really support it and it returns 0.
    The trace() function return a color anyway, not a distance..
    Other renderer that support ray tracing or newer versions of PRMan may support the function.

    Maybe some Ambient Occlusion calculation implementation in a shader might be handling distances between a point and nearby primitives directly.. but I'm not really sure.

    Anyhow, this sort of operations are probably going to require an underlying ray-tracing engine.
    Otherwise, inside a shader, when shading a point, the most you can know is the difference in position to the next point.
    The fact is that shaders are designed to deal with coherent data (one surface at the time) ..that's what makes "fast" rendering possible 8)

    Good luck !

  5. Thank you Davide!
    You are very kind, actually, I haven't found much help reading about this matter and, that's why I dared ask for your help, as a last resort. I'll try harder, maybe reading many times the RM specs would help me better understand the whole thing. I just know it must be a way :).
    Thanks a lot Davide!
    Good luck!

  6. I have recently talked to Santa Claus and, he claims the Nvidia GTX295 with it's 480 CUDA cores might be a good option. It is well-known it's a quite hungry board, it needs about 280-300 Watts but it makes sense if you think it can do about roughly 1 TFlops/s 8)
    Some Knoppix4Cuda might help see the trick :)
    The N-Body show must go on!
    Best regards Sir Davide Pasca!

  7. is the geometry static or dynamic? if its static maybe some offline pre-processing step could be used to calculate the data that you need?

    sorry i dont know much about this point rendering =)

    for me its still all about triangles.

  8. Mr. mmmore,

    I think that in this case a book like Advanced RenderMan would be more useful.

    Also tutorials from older Siggraphs, like RenderMan for Everyone in 2006.


  9. Thank you Davide for your help!
    Merry Christmas and A Hapy New Year!!!

  10. Thanksss.. nice holidays to you too !!