Nvidia Demonstrates Interactive Ray-tracing
SIGGRAPH 2010, the annual computer graphics trade show of the ACM, started is exhibition Tuesday and Nvidia demonstrated at its booth the latest Quadro Fermi graphics cards.
Nvidia demonstrated the new release of the Application Acceleration Engine, AXE 2.0. The new version is optimized to run on the new Quadro Fermi cards, but will run on any G92 or later generation cards. These engines allow for interactive ray-tracing where the user can interact with the ray-traced scene quickly without an extensive wait for re-rendering.
Nvidia’s 'mental images' division also demonstrated its iRay application, which takes the above capabilities and adds physically correct global illumination to the features. Global illumination is critical for correct lighting in most 3D digital content creation and the ability to view and interact with a globally illuminated scene will greatly accelerate the ability of artists to work quickly. In addition iRay supports network distributed rendering across GPUs.
GPU-rendered image from Bunkspeed SHOT using iray
Nvidia partners like The Foundry, RTT, and Bunkspeed also demonstrated their applications making use of CUDA-accelerated interactive ray-tracing. RTT’s application actually calculated airflow over the model and rendered the model, interactively, both using CUDA, while Bunkspeed’s SHOT uses mental images' iRay to provide interative ray tracing and global illumination.
In a few GPU generations, these can be expected to proceed from ‘interactive’ to ‘real-time’ and we should start seeing the use of real-time ray-tracing and global illumination in games.

This really does make "interactive" ray-tracing possible for the first time on a desktop... awesome. This is the sort of application the Fermi architecture really excels at. The emphasis Nvidia places on this market is probably the main reason I'll be going with a Fermi based solution for my next build, and not an Evergreen.
This really does make "interactive" ray-tracing possible for the first time on a desktop... awesome. This is the sort of application the Fermi architecture really excels at. The emphasis Nvidia places on this market is probably the main reason I'll be going with a Fermi based solution for my next build, and not an Evergreen.
this is good news. so are there 3d modeling applications that makes use of the gpu? maya/3ds?
i7 970 extreme (the old quad core one) = 48 Gflops, these video cards = 1000 to 2000 Gflops.
pwned.
any demos of real time raytracing? (20, 30fps, reasonable res ect)
obviously images like the above which look REAL aren't gonna be real time yet, but some form of awesome raytracing might be.
The problem with ray tracing isn't computational power. It's memory access. Every reflection takes only ~10 arithmetic operations to compute and then you need to find next intersection ant this is where problem starts.For a 1m triangle scene you'll have to do at least 30 reads from your data structure or even 1000s in worse case scenarios. Every read from RAM on a graphics card takes hundreds of clock cycles also the reads are random and not sequential so it won't hit very small caches and you get x16 RAM slowdown for misalignment. On a CPU you get megabytes of cache and no memory access restrictions. These algorithms work in O(n) and O(n lg n) times so you need to feed about as much data to the GPU as it can compute so for a 1Tflop chip you should connect memory that can do 1Tflop * 4 bytes = 4TB/s memory @ random reads... this can be only done with on-chip cache and for a 1m triangle scene you will need ~ 1000000triangles * 3 vertexes * 12bytes/vertex + 1 to 8 million nodes of data structure * (8 children + 1 parent) *4 bytes = way over 60MB of cache and this probably will have to be accessed by hundreds of PUs simultaneously... so it should be kept in dozens of copies.
very true
thats why triangles are lame. they are for lame rasterizers. all shapes in raytracing should be formed from much more complex geometric objects
for instance: terrain: should just a the equation for the plasma fractal you would otherwise use to generate a hight map. and bam! like 10s of bytes for your whole terrain! pwned!
I mean, the ray collision equation might be a *bit* more complex.... but Tflops!!
like how they are trying to overcome the same memory limitations in realtime rendering with tessellation.
-No car shadow reflected in body
-No road and shadow reflections on the rims
-No reflection of the mirror in the windows
-No reflection of the scenery in the glass
-No double or triple reflections
Only reflections that wouldn't be done in a super easy traditional way:
-Mirror in the bodywork (but could be done with some effort)
-Scenery reflection in the lamp shadowed by the lamp (could be done)
-Shadow on the brakes (could use a dedicated shadow here)
Only thing that looks great here is the scenery but it's only because of the detail not ray tracing. It could be rendered using Oblivion level shaders..
Good job Nvidia and I hope ATI is working on something like this too. Nothing like some good competition to push technology forward.
I hope so to,ATI's next set of cards are do out later this year hopefully they have something that can beat Nvidia.
good work