pbvh.ogg

Streaming parallel BVH construction on a dual intel X5650 (2.66gHz). Top levels of the dragon dataset (871k triangles) are build by using a space filling curve, bottom levels are constructed by SAH. The approach is more than 3x faster than nvidia´s HLBVH builder (Pantaleoni et al.) and scales pretty well on NUMA-like architectures. This enables the visualization of complex dynamic scenes (more than 1.2 Mio polygons) in less than 20ms on current
dual CPU systems.

Category:

Science & Technology

Tags:

• 3D
• Visualization
• Ray
• Tracing
• BVH
• construction

[snganRT]

@davemc0 ok, that is pretty fast. i am curios to know what the hlbvh2 really implements and for shure the results. btw have u seen my new node parallel builder constructing the boeing model ?

[davemc0]

HLBVH2 is algorithmically 5x-10x faster, not 40-70%.

1-3x/5x=40%; 1-3x/10x=70% slowdown.

In addition to the algorithmic difference, Fermi is 2x - 4x faster than the GTX280 for ray tracing; and this is 1 GPU vs. 2 CPU.

The main point, though, is nice work. I like your results. I look forward to the dual Sandy Bridge results.

[snganRT]

@davemc0 i am not allowed to publish results from a new dual socket sandy-bridge machine but believe me: 40-70% will not be enough. so i think that even this hlbvh2 thingy will be easily beaten by this algo.

[snganRT]

oohhh youtube !!! sorry davemc0

>>Hey, this looks really good for a CPU. Nice job!

So if this is three times faster than Pantoleoni's HPG 2010 paper that would make it 40% to 70% slower than Garanzha's HPG 2011 HLBVH2 paper. And another 2X slower when comparing a single CPU to a single GPU.

Some approaches use space-filling curve for top level and SAH for bottom. Others do the opposite. How did you pick your approach?<<