News for hackers. Stuff that matters.
Re:Tensorflow 2.0.0-Alpha0 by izacus on Wednesday March 06, 2019 @03:52PM (#19322650)
> Yes, real-time ray tracing, the new feature in these cards, is so bad that it's barely usable — but we're used to that, v2 will be better. I have a lot of sympathy for NVIDIA engineers trying to improve the state of the art, it's never easy.

Can you explain this more? All the tests of RTX in actual games have been very positive and showed significant improvements in visual quality.

Re:Tensorflow 2.0.0-Alpha0 by dragontamer on Wednesday March 06, 2019 @04:00PM (#19322728)
The results I've seen show the $1200 GPU playing at 1080p resolution with less than 60FPS.

RTX is clearly a compute-hog: its barely usable even if you pay for the best-of-the-best GPUs. I mean, not to knock Nvidia down or anything, raytracing is one of the most computationally difficult problems in existence right now.

But from a practical and pragmatic perspective, you suffer a major loss in frame-rates and force the GPU to drop in resolution to make the jump to raytracing. Even if you spent $1200 on the card...

-----------

I'm personally excited to see offline renderers use the RTX features to accelerate offline raytracing. That's probably the more important use of the technology. As it is, RTX isn't quite fast enough for "real time" yet. Just grossly accelerated offline raytracing (which is still impressive)

Re:Tensorflow 2.0.0-Alpha0 by izacus on Wednesday March 06, 2019 @05:29PM (#19323581)
According to PC Gamer and Digital Foundry, the 2080Ti can drive Metro Exodus with RTX on at average 55fps with drops down to 30fps at 4K resolution so your claim seems grossly exaggerated - https://cdn.mos.cms.futurecdn.net/YhDHpgGrAmpmnP4LUBEgvg.png

Also last I read, RTX really isn't designed for offline raytracing and doesn't really bring much to the table. Its use is in realtime.

Re:Tensorflow 2.0.0-Alpha0 by dragontamer on Wednesday March 06, 2019 @06:47PM (#19324186)
https://www.hardocp.com/article/2019/01/07/battlefield_v_nvi...

Battlefield V Ultra 1080p RTX on is ~63 FPS average (not minimum FPS, but average), which means it will regularly dip below 60 in practice.

> Also last I read, RTX really isn't designed for offline raytracing and doesn't really bring much to the table. Its use is in realtime.

On the contrary! RTX Cores are a hardware-accelerated BVH-traversal. That has HUGE implications for the offline rendering scene.

See NVidia Optix for details: https://devblogs.nvidia.com/nvidia-optix-ray-tracing-powered...

IMO, this is the killer-feature of RTX. Accelerating those Hollywood Renders from hours-per-frame to minutes-per-frame. NVidia Optix takes industry-standard scene trees and can use the RTX Cores to traverse them for hardware-accelerated raytracing.

Or more specifically: coarse AABB Bounds checking, which is a very compute & memory heavy portion of the Raytracing algorithm.

Even if RTX is too slow for video games, any improvement to offline rendering is a huge advantage.

Re:Tensorflow 2.0.0-Alpha0 by ykombinaught on Tuesday March 12, 2019 @11:06AM (#19368827)
Optix was developed a decade ago and doesn’t require RTX cards. It run on any Maxwell or newer card. It’s NVIDIA’s version of RadeonRays.

What NVIDIA is trying to sell people is a denoiser.

And are you sure RT Cores even exist or are they just tensorcores by a different name?

https://youtu.be/3BOUAkJxJac

Re:Tensorflow 2.0.0-Alpha0 by dkersten on Thursday March 07, 2019 @12:07PM (#19329749)
In the RTX announcement speech, they repeatedly mentioned the savings for offline rendering (taking fewer computers less time and fewer watts to render the same quality). I came away from watching that thinking that the card was primarily positioned for offline rendering.
Re:Tensorflow 2.0.0-Alpha0 by rolleiflex on Wednesday March 06, 2019 @04:01PM (#19322735)
My impression from the analyses I've seen so far is that it barely improves the visual quality, since the game designers learned to fake it so well over the last two decades, and it comes at a significant cost to frame rate. In particular, good ray-tracing with acceptable noise levels comes at around 500 rays a point. RTX cards are capable of 2 rays a point.

Here is a comparison of the regular global illumination vs RTX real-time raytracing, particularly take look at noise levels generated by RTX in the background. https://www.youtube.com/watch?v=CuoER1DwYLY

Re:Tensorflow 2.0.0-Alpha0 by dragontamer on Wednesday March 06, 2019 @04:12PM (#19322832)
>In particular, good ray-tracing with acceptable noise levels comes at around 500 rays a point. RTX cards are capable of 2 rays a point.

NVidia has that temporal denoising algorithm though. So noise is smoothed between space and time. When you have 60-frames per second, you can "recycle" old frame's raytracing data for the current frame.

As long as the raytracing lights rarely change, it'd work out pretty well. The fast-moving lights would probably have to still use rasterization techniques, but the demos of the temporal-denoising algorithm from NVidia are very impressive.

FYI: Blender Cycle's denoising is still kinda bad even at 1000 or even 5000 samples (at least, when you look at low-light areas close to walls). You gotta zoom in and look, but artistic CGI requires many, many samples before raytracing looks smooth. Its a major advancement for NVidia to come up with a raytracing denoising algorithm that works well for video games on only 1-ish samples per pixel.