Ray-Tracing, meaningful performance metrics and alternatives? *spawn*

Discussion in 'Rendering Technology and APIs' started by Scott_Arm, Aug 21, 2018.

  1. pharma

    Veteran Regular

    Joined:
    Mar 29, 2004
    Messages:
    2,485
    Likes Received:
    1,237
    Introduction to Real-Time Ray Tracing with Vulkan
    October 10, 2018
    https://devblogs.nvidia.com/vulkan-raytracing/
     
    #241 pharma, Oct 10, 2018
    Last edited: Oct 10, 2018
  2. Ext3h

    Regular Newcomer

    Joined:
    Sep 4, 2015
    Messages:
    312
    Likes Received:
    256
    Some new performance numbers on the "Speed of Light" demo from talks on this years GTC:

    It reached 24fps ("We define real-time as 24fps"). At FHD with 20% super sampling. Running on not one, but two RTX 6000 in SFR.

    Effectively, they got around 15 (hitting / missing) rays per pixel per frame. Not than fancy "10 Giga-rays per second" claimed by Jensen, but a far more modest 300 Mega-rays per second and card against a non-empty BVH, and that despite 95% of the GPU load being attributed to just the "hidden" parts of the ray tracing, not actual shading. With that performance number being relatively stable within a broad range of model complexity.

    15 rays per pixel is a joke, for anything but the most trivial, pure specular materials or pure shadow calculation. For a couple of effects, like scattering with transparent materials, they still had to fall back to screen space post processing effects, partially because unsupported by the card, partially because that requires significantly more rays. Also no geometry in the scene other than the car and the box lights, in order to waste as little rays outside the relevant area as possible.

    Transfer these numbers to a modern 4k @ 60fps requirement, and you are left with less than 1 ray per pixel per frame, after already using a hybrid tracer. Not even enough to cast shadows against a single point-light.


    The "dancing robot" demo from GTC? Ran on 4 GPUs, "live". Probably with the same "cinematic restrictions".
    Yes, it's still impressive. But it's about factor 16-32x away from the efficiency and performance required to make this in any form relevant for actual real-time computer graphics.

    And keep in mind that even Jensen acknowledged this time that Moore's Law is dead, even for Nvidia. They can still afford to add more special function units (not running in parallel) with smaller nodes, but they are hitting a wall with power density as well, and the only way around that wall would be to switch from high performance to a low power node, getting another factor 2. Or in other terms: That required performance increase is simply not possible.
     
    Lightman, Silent_Buddha, Malo and 2 others like this.
  3. chris1515

    Veteran Regular

    Joined:
    Jul 24, 2005
    Messages:
    3,051
    Likes Received:
    1,573
    Do you have a link to this performance?
     
    pharma likes this.
  4. Jupiter

    Veteran Newcomer

    Joined:
    Feb 24, 2015
    Messages:
    1,288
    Likes Received:
    808
    As a beam one counts every single request of intersection with the scene. So something like transparency are "N" rays depending on how deep one wants to go. The 10 Giga-rays per second are measured under ideal conditions and in practice one has clearly less. The latencies for texture accesses, complex shaders with more registers etc. are quickly pushing it down (this is normal). Practically the performance of raytracing depends on how many near misses one has, how divergent one is and what additional latencies one gets. I wouldn't give too much about the value but look for the papers and applications that show it in the context of real content. It only needs the value to compare the SKUs and that's what Nvidia raytracing experts say.

    For example "Ray-Tracing Operations Per Second" is marketing and one will see what is established as a measure when other manufacturers also have something. Then it will run out on certain benchmark scores as usual.
     
    #244 Jupiter, Oct 12, 2018
    Last edited: Oct 12, 2018
    pharma, Lightman, OCASM and 1 other person like this.
  5. Ext3h

    Regular Newcomer

    Joined:
    Sep 4, 2015
    Messages:
    312
    Likes Received:
    256
    The slides are not online yet. Check "Deconstructing the 'Speed of Light' - Porsche, NVIDIA RTX and Unreal Engine 4 (Part 1) at GTC Europe." once they are.
     
    Lightman, BRiT and chris1515 like this.
  6. chris1515

    Veteran Regular

    Joined:
    Jul 24, 2005
    Messages:
    3,051
    Likes Received:
    1,573
    thanks
     
  7. OCASM

    Regular Newcomer

    Joined:
    Nov 12, 2016
    Messages:
    502
    Likes Received:
    439
    The solution is simple: run ray tracing at less than 1spp and use denoising and upscaling to compensate. Even under those contraints it's far better than rasterization.

    Also, just render at 1080p and upscale to 4K. Only the guys at Digital Foundry would notice.

    @sebbi agrees:

     
    #247 OCASM, Oct 12, 2018
    Last edited: Oct 12, 2018
    Jupiter and pharma like this.
  8. milk

    Veteran Regular

    Joined:
    Jun 6, 2012
    Messages:
    2,276
    Likes Received:
    1,636
    So much for Ray Tracing spelling the end of tricks and hacks.

    It just works, but not really.
     
    Lightman, Kej, Silent_Buddha and 3 others like this.
  9. jlippo

    Veteran Regular

    Joined:
    Oct 7, 2004
    Messages:
    1,166
    Likes Received:
    210
    Location:
    Finland
    At least it's new set of hacks.
    Or different set of old hacks. ;)
     
    London-boy, pharma and OCASM like this.
  10. OCASM

    Regular Newcomer

    Joined:
    Nov 12, 2016
    Messages:
    502
    Likes Received:
    439
    The end of the old tricks and hacks :cool:
     
  11. pharma

    Veteran Regular

    Joined:
    Mar 29, 2004
    Messages:
    2,485
    Likes Received:
    1,237
    NVIDIA RTX Effects Cost 9.2 ms in Remedy’s Northlight Engine, Running at 1080p on an RTX 2080 Ti

    https://wccftech.com/nvidia-rtx-9-2-ms-remedy-northlight/
     
    milk and OCASM like this.
  12. pharma

    Veteran Regular

    Joined:
    Mar 29, 2004
    Messages:
    2,485
    Likes Received:
    1,237
    October 16, 2018
     
    OCASM likes this.
  13. A1xLLcqAgt0qc2RyMz0y

    Regular

    Joined:
    Feb 6, 2010
    Messages:
    932
    Likes Received:
    212
    OCASM and pharma like this.
  14. CarstenS

    Veteran Subscriber

    Joined:
    May 31, 2002
    Messages:
    4,720
    Likes Received:
    1,939
    Location:
    Germany
    Also, pretty meaningless numbers without the time required by the default implementatinon without raytracing. If those effects or their equivalents were at, say 3 ms, the judgement of RT performance would be quite different than if the raster-based effects would need 12 ms. Of course, those are just made-up numbers, just to illustrate a point.
     
    milk, OCASM, jlippo and 2 others like this.

Share This Page

  • About Us

    Beyond3D has been around for over a decade and prides itself on being the best place on the web for in-depth, technically-driven discussion and analysis of 3D graphics hardware. If you love pixels and transistors, you've come to the right place!

    Beyond3D is proudly published by GPU Tools Ltd.
Loading...