Recent content by keldor

My understanding of scientific HPC algorithms is that at their hearts they very often rely on huge amounts of matrix multiplication. This is not something restricted to neural nets by any means. Since the Tensor Cores are in fact specialized matrix multiplication hardware, and since they added...

Addressing the topic of the thread more directly, games will end up requiring ray tracing. The reason is that as long as it isn't a requirement, the developers will have to maintain two very different render paths, which is that much extra work for both the programmers and the artists. Pure...

The difference between hybrid ray tracing and pure raytracing is a red herring. Turing is perfectly capable of rendering with pure raytracing, but the question is why would you want to? For primary rays, rasterization is more efficient and does the exact same thing. The only think ray tracing...

Remember, it's splitting a big register bank between some number of wavefronts/threads. This means the number of registers to a wavefront needs to divide the number of registers in the register bank or else there will be left over ones that are just wasted. As for asynchronous compute, this is...

I wouldn't read too much too much into the memory speed numbers. It's basically overclocking. 4400 is the completely stable value that they feel comfortable making the baseline, and 5100 or 5133 are the maximum levels they achieved, but may not be reachable on every chip, and may not be...

That's pretty spectacular. Granted, there's room for improvement in the pure rasterized version (adding in a bit of fixed ambient would help), but the amount of dynamic geometry in what we see makes baking ineffective.

Looks like video compression artifacts to me. I can see them all over the place. The bitrate leaves something to be desired.

The API is very high level - shaders dispatch some rays, and at some point a hit/miss/whatever kernel is executed with the results. There's a lot of resemblance to Nvidia's dynamic parallelism in Cuda, honestly. Rays are already inherently batched to some extent at the wavefront level...

All I see is a bunch of PR drivel. And comparing their product due sometime in the next few years with the programming model of GPUs about a decade ago. I'm not impressed.

There's something wrong here - 4 GB for 35 M polys equates to over 100 bytes per poly! Is this model being stored as plaintext or something?

Just glancing at that paper it appears that the performance of the algorithm completely collapses for any reasonably large triangle count. Going from 30 FPS on very low poly scenes (only a couple thousand triangles) to less than 1 FPS for medium poly scenes (a few hundred thousand triangles)...

SMP could be used for foveated rendering, but it wouldn't be all that effective. The problem is that your texels (for lack of a better term) will end up all stretched out, so they'll be much different densities in different directions. You can get a linear reduction in pixel count per DPI with...

The problem is that with shadow maps, you basically have to render a depth cubemap (if you want the light to cast shadows in all directions - sunlight doesn't need this) for every single light source. There really just isn't any good way to do this while maintaining both quality and speed, not...

It means you can bias your sampling to get better accuracy in the section of the screen your eye is looking at. On the other hand, undersampling tends to lead toward flickering type artifacts, which your peripheral vision is actually more sensitive to than your center of gaze. Experimentation...

I mean, I'd argue that if your chromatic aberration filter is so wide that the color banding is visible, you're abusing the feature. :wink: Simulating low quality camera optics has always baffled me. I mean, it makes sense in a movie where you're compositing in with video from a real physical...