Hacker News Reader

Comment by cubefox | Hacker News Reader

The closest guy to Carmack today seems to be Brian Caris, who developed virtualized geometry[1], a rasterization based rendering technique where the geometric detail of static objects is adjusted to the screen resolution in real time. There is another talk (can't find it right now) where he mentions Carmack as an inspiration. I wonder what Carmack would be working on today if he was still in the game engine business.

[1] https://advances.realtimerendering.com/s2021/Karis_Nanite_SI...

View on HN

He was a little bit. I wonder what he was up at meta before switching to agi

VR headsets.

Who made DLSS? Using ML to denoise ray tracing is massive.

I think DLSS doesn't do denoising. DLSS uses an ML algorithm to upscale frames using additional information like z buffer and motion vectors. Denoising is used to drastically reduce the number of required rays for ray tracing. I would guess denoising is done first, using a non-ML algorithm, and DLSS is done afterwards.

DLSS does the denoising since DLSS 1.0 see the DLSS announcement https://www.youtube.com/live/4_g_Y0W1Xn8?feature=share (pardon the crappy gaming channel, it looks like Nvidia weren’t running their own YouTube the time)

It adds frames since I think 3.0.

That's odd, since you can use ray tracing without DLSS. I'm very sure it uses denoising even then.

There may be some confusion between upscaling and denoising. You might notice I keep editing my comment because it’s really really hard to find old Nvidia release announcements.

Update: https://youtu.be/6O2B9BZiZjQ looks good around 3:54 Nvidia refer to “deep learning for image denoising” which to me seems like DLSS.

Ray tracing without DLSS might be a different denoising technique. Is it really slow when you turn it on ?

There's some really big innovations happening in e.g. rendering like that; I wouldn't be surprised if the complexity and work involved in just Nanite is 10x that of the whole Quake engine.

Similarly there's the work nvidia is doing, using AI technology to upscale graphics instead of rendering things at 4K.

Hobbyists have already made their own implementation of Nanite. It's not that complicated.

It has been over a year since the algorithm has been explained in quite a bit of detail.

Hobbyists make complicated things quite regularly, although I don't know about 10x quake.

10x a quake for modern hardware, maybe (could be quite simple)

10x software rendered dos quake with loads of asm, probably not.

In that sense, neither is the quake engine.

The idea is that, the derivative came after the virtuoso. Of course what came after is more complex costs more takes more is more impressive blah blah. Dismissing those that thought the thought or solved for a problem before it was, is such a dumb thing people do.

Reply on news.ycombinator.com