Assuming that the GPU will not be used for any other graphics operations, my fair guess is that the economics involved in developing a GPU solely for this purpose "may" not always end up favourably . The expenditure associated with design, manufacture and testing of an IC can only be justified through economies of scale.

Obviously the speed and efficiency is going to be much better than a general purpose programmable GPU pipeline.

A more practical solution is to use an FPGA, which are cheaper, available right now, and programmable with OpenCL.

Good info from you both. Thank you. I agree with you Manmohan, that its very expensive to build an entire new kind of GPU, and surely needs volyme to make sense from an economic standpoint. But, if its possible to engineer a dedicated GPU for pathtracing that infact -would- generate an enormous performance boost, making it possible to run it in realtime on higher resolution/sufficient samples, the practical implementation is obvious and the economic returns are massive.

Lubomir, would a FPGA more or less give the same result as a non purpose GPU or do you think it could be much faster? This is not my field so forgive my ignorance. Any tips are welcomed, even those who seem obvious.

Here is test running on 2 Nvidia cards:

http://www.youtube.com/watch?v=evfXAUm8D6k

Imagine if it was possible to create a chip that could push this in 30-60 fps in good quality. It of course, will be one day, im just curious if it can be done sooner than one might think.

I agree that an specific architecture will perform better, but the cost of developing a chip focused on path tracing will probably result too high. GPU makers obtain most part of their revenue from mainstream computers and laptops, thereby financing the high engineering costs of graphics hardware.

As far as I know, one of the limiting factors of path tracing is the irregular code, with a significant amount of branches and thread divergence. According to NVIDIA's documentation, the next generation of Kepler GPUs will have a feature called Dynamic Parallelism which may be useful when trying to develop efficient implementations for path tracing. You can find more information in the "Kepler GK110 whitepaper" (the pdf can be easily found in Google) and some preview examples in the CUDA 5.0 SDK beta release.

Unfortunately, this feature may only be available in the high-end Tesla K20 series. It is very likely that it will not be supported in the normal GeForce cards, at least for some time after the official GK110 launch (around Q4 2012).

Hi Michael,

As I see the alghoritms must be adpted for the new graphics cards, like Kepler architecture, I think if you Nvidia Tesla K20 for calculations and a Quadro (not Kepler yet) for rendering you will have a great performance. Note that this new Tesla will accept up to 32 kernels running concurrently and the dynamic paralelism will avoid the CPU to get busy transfering the kernels results to CPU, it will only return the final result to the CPU. So I say that software must adapt its algorithm for the new architecture. And there are other new features on this new architecture that the industry will benefit. I dont think building a new GPU for only this porpouse should be done since we have great cards every year, but as Lubomir said using FPGA is an idea.

I don't have direct experience with FPGAs and speed data seems hard to find with a quick Google search :). My understanding is that you should be able to get better speeds with FPGAs than with GPUs, since you are "hard-coding" the algorithm.

BTW, to my understanding, Dynamic Parallelism will only reduce the need to communicate with the CPU by launching sub-grids to the original problem at run-time. The GPU architecture does not change, i.e. all cores in a Compute Unit will still execute in lock-step, so branches will be just as expensive. And K20 is the only Kepler that will have the feature.