GPU Ray Tracing Performance Comparisons [2021-2022]

Not sure where to put this.

New Research paper from Nvidia that shows AI accelerated ray tracing
We propose neural control variates (NCV) for unbiased variance reduction in parametric Monte Carlo integration. So far, the core challenge of applying the method of control variates has been finding a good approximation of the integrand that is cheap to integrate. We show that a set of neural networks can face that challenge: a normalizing flow that approximates the shape of the integrand and another neural network that infers the solution of the integral equation. We also propose to leverage a neural importance sampler to estimate the difference between the original integrand and the learned control variate. To optimize the resulting parametric estimator, we derive a theoretically optimal, variance-minimizing loss function, and propose an alternative, composite loss for stable online training in practice. When applied to light transport simulation, neural control variates are capable of matching the state-of-the-art performance of other unbiased approaches, while providing means to develop more performant, practical solutions. Specifically, we show that the learned light-field approximation is of sufficient quality for high-order bounces, allowing us to omit the error correction and thereby dramatically reduce the noise at the cost of negligible visible bias.
Source:
https://research.nvidia.com/publication/2020-11_neural-control-variates
Youtube video:

Extremely smart :idea:
 
Fast Parallel Construction of High-Quality Bounding Volume Hierarchies - Nvidia Research

4.3 Dynamic Programming
While the naive algorithm is straightforward, it is also woefully inefficient. For instance, n = 7 results in a total of 1.15 million recursive function calls and an even larger number of temporary solutions that are immediately discarded afterwards. To transform the algorithm into a more efficient form that produces an identical result, we make three important modifications to it:
• Remove the recursion and perform the computation in a predetermined order instead.
• Represent S and P as bitmasks, where each bit indicates whether the corresponding leaf is included in the set.
• Memoize the optimal solution for each subset, using the bitmasks as array indices

It would be interesting to see what DPX could give benefit to this..
 
I doubt NVidia has waited 9 years to put this research into practice.
The TRBVH (technique described in the paper) were actually used in the OptiX until version 6.x but I don't know it's still used in the 7.x since the API has changed entirely to support hardware raytracing.
Now we know Nvidia has a patent about hardware SBVH traversal, and they have DPX instructions that can (potentially) accelerates TRBVH construction speed. Even if the performance is not good enough for the games they can bring back it for OptiX anyway.
 
The TRBVH (technique described in the paper) were actually used in the OptiX until version 6.x but I don't know it's still used in the 7.x since the API has changed entirely to support hardware raytracing.
Now we know Nvidia has a patent about hardware SBVH traversal, and they have DPX instructions that can (potentially) accelerates TRBVH construction speed. Even if the performance is not good enough for the games they can bring back it for OptiX anyway.
Optix 7 dates from 2019 - the bitwise instructions in Turing look like candidates for the implementation of these techniques - or at the very least, some of them.

So what does DPX bring that's not in Turing that's covered in TRBVH?

9 years is a VERY LONG TIME in computing.
 
Back
Top