Hacker News Reader

Comment by spyder | Hacker News Reader

I believe Magic Leap has a new kind of display technology, but the article is pretty bullshit-sounding with statements like this:

"Throw out your PC, your laptop and your mobile phone, because the computing power you need will be in your glasses, and they can make a display appear anywhere, at any size you like."

Yeah sure, they not only made a new display tech but integrated the computing power of desktops or laptops into glasses. I wonder how they cool the GTX 1080 cards on your head...

View on HN

Uploadvr demystified this stuff a couple years ago:

http://uploadvr.com/magic-leap-how-it-works/

Magic leap is a pair of glasses you wear on your face that displays AR imagery through a clear glass lens. There's a cable leading from these glasses to your pocket where you have a tablet-size device doing all the heavy lifting, most likely using mobile SoC's. Pretty sure you can't run a desktop CPU and GPU on a pocketable device.

You don't get 1080gtx performance or anything remotely close to that with a setup like this. Its mobile-level graphics. Oculus's recent showing of its mobile SoC-based Santa Cruz prototype shows fairly poor graphics compared to PC-based VR like the Rift or Vive. I imagine the Magic Leap has less pixels to push considering it isn't doing any backgrounds, but still, those very impressive marketing shots of high resolution dragons flying around and interacting with you might not really be possible without a lot of compromises on graphic quality, fov, framerate, etc. Worse, no one wants to talk about transparency, especially in well-lit rooms. The few hands-on Hololens reports we've gotten make this out to be a big issue, as well as fov, and fundamentally the Magic Leap and what hololens is using could be the same or very similar technologies.

My gut feeling is that no one has yet to handle the transparency issue (and this is why MS can show off the hololens with confidence) and because of that it makes sense for Magic Leap to keep its product secretive. The tech press was not too kind to Hololens prototypes due to transparecy issues and Magic Leap doesn't need that kind of negative press as it continues to fund-raise and develop its product.

How many years will it be until Pascal-like performance is available in a tablet form-factor?

Maybe never. Pascal is 16nm process and the mid-range gaming card, the 1070, uses up to 150 watts. A "powerful" mobile chipset uses a tiny fraction of that even at its highest settings. 100 watts for the CPU and another 150watts for the GPU is a staggering 250watts of continuous power no pocket-able lithium-ion battery can handle.

If we want current gen graphics we're more than likely looking at a PC backpack solution. More than likely we'll probably be looking at some low watt solution, like the 35W AMD chipset in the Solus Q and accept a lower level of graphic quality, but still well above mobile standards. This may involve wearing a battery belt or battery hip-pouch for extended use. For lower run times the battery can be part of the headstrap with no need for anything to be pocketable.

Part of me says just distribute the battery all around, thighs, legs, torso and you'll at least have good power. Another part of me says this is ludicrous, and is akin to strapping thermite and a bunch of loose strike anywhere matches into pouches all around your body.

For many professional applications it is not hard to imagine people wearing some kind of vest or belt related to this.

Only by some combination of streaming and prediction; doing the rendering in the cloud, and using clever technology to hide latency.

Which Pascal? Assuming the smallest desktop Pascal, GP107 is at 768 CUDA cores. Tegra X2 Pascal has around 256 CUDA cores. They seem to be clocked similarly. The Tegra also has a lot lower memory bandwidth and stuff like that but we'll put that aside for now. We'll need the Tegra to x3 in performance. Maybe in 6? years we can get a GP107 performance in tablet form factor. The GTX 1080 has 2560 CUDA cores. That's another 3x from GP107. So maybe 12 years for ~1080 performance in tablets.

One important differentiating factor between this and other (consumer) HMD's is the scanning display.

The Oculus et al need to render a rectangular matrix of evenly spaced pixels. If you want a larger field of view or higher resolution this comes with an exponential increase in number of pixels. With a high frequency scanning display and eye tracking you are not locked to a static spatial resolution (see https://en.m.wikipedia.org/wiki/Foveated_imaging).

There's a whole bunch of other problems when it comes to a render pipeline for something links this, but at least on the physical front it enables super high perceived resolution with less overall points of light to render.

Foveated rendering only became relevant when Nvidia recently started supporting Multi-Resolution Shading. As new techniques and tools like this get better support we can be a lot more efficient with lower powered hardware.

It was interesting to hear Carmack last year talk about re-introducing interlacing as a way to reduce latency.

Although, I'm just reinforcing what you're saying =)

Huh, interlacing is a neat idea to reduce latency, but I'm guessing you'd want a form of intra-frame temporal anti-aliasing, otherwise it might look pretty fugly. Whether or not this would neutralise any latency advantage due to the computational overhead of doing this, I'm not sure...

Found where Carmack talked about it[1].

For VR, the consensus I've seen seem to push for multi-sample AA > full-screen AA > temporal AA (I feel like I've even seen no AA is better than temporal AA). I'm not quite sure if it's performance, architecture (they really prefer forward renderers instead of deferred renderers for low latency), or aesthetics. In the little bit I've done, when just playing with knobs temporal AA looks better to me. Without AA specular highlights are way too distracting and pop too much.

This reference[2] talks about the extra velocity buffer needed for temporal AA and the fact it tends to over-blur which can fuzz out fine details (low resolution is a touchy subject in VR).

There's a lot of moving pieces when making decisions. In talking about interlacing, he was ideally talking about not having to create a whole image buffer before sending data to the hardware--just rendering parts (well, scanline) of the image that moved.

[1] https://youtu.be/gn8m5d74fk8?t=10m54s [2] https://www.youtube.com/watch?v=Xk3WUk5T2TQ&feature=youtu.be...

>If you want a larger field of view or higher resolution this comes with an exponential increase in number of pixels.

How does that work? Shouldn't it be just linear in FOV? Or quadratic if you are talking about solid angle? This doesn't seem like one of those things that grows exponentially.

From what I read here, I think they are aiming for a system that has 'enough' resolution on the entire field of view.

'Enough' is retina resolution, but only on the fovea, and way less the further away you go from the fovea. That can significantly decrease the amount of work needed to render a scene.

Moreover, color-wise, you don't need color except for the center of your field of view (you need cones to see color, and cones can only be found in the fovea).

Building such a screen probably isn't that hard; you would need to build new production lines because everything out there is built for rectangular grids of equal-sized pixels, but that probably is 'just' a matter of investing money and time.

Writing the software to efficiently draw scenes on such a screen probably is a bit harder; there will be lots of optimizations that need to be discovered (what's the equivalent of Bresenham line drawing, for example?)

The biggest challenge will be to keep the high-resolution part of your screen centered on where the retina looks. That requires eye tracking and some way to move your screen. Saccades move at hundreds of degrees per second; you would have to match that.

If that's what they are building, and they succeed, I think it will take over the world, even if they initially sell it at $5000 an eye.

Reminds me that a company a while back tried to pitch such a system on the cheap. It was basically a bulky visor that you put your smartphone into. inside the visor was a angled piece of glass that would make your phone screen appear to be hovering in the air in front of you.

An AR Cardboard would be awesome. I searched some weeks ago if someone worked on this without luck. Do you remember the name of this company?

http://www.caputer.com/

- Unlike VR, they don't have to render backgrounds

- FOV is apparently their big advantage, along with focus zones

- They could do hard occlusion with an LCD mask. Or they could punt til v2, and soft-occlude with blur using their stacked diffraction plates.

> - They could do hard occlusion with an LCD mask. Or they could punt til v2, and soft-occlude with blur using their stacked diffraction plates.

Unfortunately, you cannot easily achieve "hard occlusion" with an LCD mask. Pinlight display is a project that did explore that idea: http://pinlights.info

LCDs can attenuate individual rays of light. However, the images we see of real objects in our environment are generated from light in the form of planar wave fronts. So, if you have a near eye LCD micro display and you turn on a single LCD pixel in an attempt to block light emanating from a certain point on a real object, most of the light emanating from that point will simply go around the LCD pixel, enter the pupil and for a point on you retina. The total brightness of the image will dim a little dependant on the size of the LCD pixel, but not by much for a single microdisplay pixel (i.e. a few micron pitch).

The Pinlight papers do talk about ways around that issue, using LCD diffraction pattern masks, but I still think the residual diffraction and reduced overall lens transparency associated with the LCD micro display makes the approach an even bigger challenge.

Depends on size of occluded area? If I block the sun with my thumb, some light does go around, but it's still a decently effective mask.

If you hold out your thumb between your eye and the sun, such that your thumb appears to just block the edges of the sun, then you will be effectively blocking the entire set of overlapping wavefronts emanating from the sun (that would have formed multiple "pixels" forming the sun image on your retina), preventing their arrival at your cornea and ultimately entering your pupil. Notice that your thumb is about about 1.5ft from your eye.

Could you emulate that with a transparent LCD display? Yes, but that display would need to be large and about 1.5ft from your eye. Not really practical for a near eye (head mounted display).

If you try to accomplish this occlusion directly with an LCD microdiplay (such as those found in many projectors) in a near eye application, at best you end up with a blackish smudge with very fuzzy edges. To see this in action, look at Figure 11.D in the following paper:

http://www.cs.unc.edu/~maimone/media/pinlights_siggraph_2014...

Thanks for the paper link. Is there a plausible way to use destructive interference to create true black? I remember patents around this.

Most of the light that is received by the retina in day to day life is highly incoherent thus it is unlikely that you can produce a destructive wavefront for it.

One of their patents mentions using stacked diffraction optical elements. Details unclear:

>[0216] Such may be used to cancel light from the planar waveguides with respect to light from the background or real world, in some respects similar to noise canceling headphones.

http://appft.uspto.gov/netacgi/nph-Parser?Sect1=PTO1&Sect2=H...

> One of their patents mentions using stacked diffraction optical elements [DOE].

> http://appft.uspto.gov/netacgi/nph-Parser?Sect1=PTO1&Sect2=H....

I believe this particular excerpt is referring to the use of the layered and controllable DOEs to curve light projected from their scanning fibers that ultimately forms the virtual object image on the retina, rather than using the DOEs to attenuate ambient light (see my reply to Kelsolaar above).

However, if I'm interpreting their literature correctly, you are correct in that they are able to use (at least a portion of) these DOE layers to generate the moire patterns that will ultimately be capable of form black/clear image masks.

Would collimator in front help?

I agree that the lack of coherence (spatial & temporal) of ambient lighy would makes the use of diffraction gratings to actively create destructive interference patterns on the retina a futile efftort. Ambient light is generally a white lighy (lacks temporal coherence) and many wavefronts originate closer to the eye than the optical infinity (~8m) and are therefore fairly divergent (curved and thus lack spatial coherence). The resulting interference patterne would therefore be very blurry. Too blurry to form a useful black pixel image on the retina.

On the other hand, I believe ML is proposing to use a gratings to display black/clear images. In this setup, the incident ambient light wavefronts are selectively redirected using bragg gratings (why they keep saying Braff gratings" in their literature is beyond me). This technique has been around for many years and is still captivating to watch even in this day and age - see:

https://www.google.com/url?sa=t&source=web&rct=j&url=https:/...

The realization that this affect retains its "sharpness" for near eye applications in concert with (I believe) the realization that you can utilize the electonically-controlled diffraction grating layers you already have in your device (for the purpose of controlling wavefront curvature to support eye depth accommodation) depth is really quite clever. It's truly an exciting prospectct!

> There's a cable leading from these glasses to your pocket where you have a tablet-size device doing all the heavy lifting, most likely using mobile SoC's. Pretty sure you can't run a desktop CPU and GPU on a pocketable device.

> You don't get 1080gtx performance or anything remotely close to that with a setup like this. Its mobile-level graphics.

A.) the pocket device is the light source array that feed the scanning fibers. A processor drives each light source linearly. Neither the light source nor the driving processor are likely to be head mounted for quite some time, but I don't see that being an issue given the dramatic improvement over existing interfaces (i.e. I'd rather have a pair of glasses with a tether than a helmet any day of the week);

B.) In light field VRDs, resolution is not limited by active matrix array addressing as it is to a degree in LCD displays of mobile devices. Instead, resolution primarily depends on how quickly you can resonate the tip of a scanning fiber and concurrently how quickly you can pulse that fiber's light source in a synchronized manner (each pulse matching the bit color/intensity of the intended pixel);

C.) Multiple fibers form a composite image;

D.) Maximum resolution is only needed in the foveal region (i.e. a few deg about the center of user's eye gaze at any given time), so eye tracking can be utilized to minimize processing. The remainder of displayed content is low resolution, albeit still critical, peripheral imagery.

My team & I have been batting around with a slightly different approach to the light field VRD, which we're thinking about simply releasing as a FOSH project.

> we're thinking about simply releasing as a FOSH project.

More details would be fantastic

Well, our light field display concept is fairly different from ML's in the manner in which we scan the light into the eye. We propose using very narrow collimated wavefronts instead of broad/potentially curved wavefronts. The system can still be used to support accomodation, potentially to a greater degree than is possible using layered DOE's (I don't know for sure) which we do not propose to use.

The resultant display has an inherently smaller exit pupil, but we'll layout a method to actively expand the exit pupil, enough to support typical maximum eye rotation range and minimum eye pupil dilation (~2mm) without vignetting the virtual image with natural eye movements. The benefit of this method is virtually unlimitted FOV.

Certainly still in development, but we think releasing what we've come up with thus far into the wild as FOSH and seeing what the crowd can bring to the design is a fun idea.

I'm currently working on a simple "Hello World" write up that will explain the device components, some basic plans for prototyping done to date and the development hurdles that still need to be overcome. I'm planning to use this to introduce the project. It's definitely not an easy project, but I don't see any component that can't be fabricated by a crafty homebrew engineer.

Fantastic, thanks. Looking forward to the hello world write up

There is a bigger problem, as well. Tracking, the glasses will need to use a lot of the GPU (and CPU) processing power for simply tracking the environment. So they don't have backgrounds, that doesn't account for much anyway, like maybe 10% of GPU power goes to that. Tracking an environment takes like an entire card at the moment.

Microsoft created an ASIC they call an HPU (holographic processing unit) that does this work for the GPU. It takes data from the depth sensors and sends the resulting point data straight to the CPU.

One of the researchers posted a 30ish min video on the hardware inside the Hololens on YouTube. It's this HPU / sensors that Microsoft is licensing to the hardware manufacturers for the upcoming VR and AR headsets. We'll know more December 8th, and after their hardware expo later this year.

To be fair, if the tracking and resolution is first-in-class, there are a billion business applications that won't require major graphics horsepower. I'm going to buy a vive soon to play around with virtual desktop, but I don't have high hopes for it since I know the resolution isn't there yet. If they can come up with something light weight and high resolution that lets me see the rest of my environment I'm all in.

Don't buy the Vive for virtual desktop unless you don't care about your money. I have the Vive and it's awesome for games. Virtual desktop is cool too, but it's not something you want to use for any length of time. It's like looking at your phone stretched out on a 50 foot screen. VR videos are about the only use. Anything else is maybe just a glimpse into the future. The resolution is too low right now.

Have you tried adjusting your supersampling settings?

Well the Oculus Rift VR quality is better than Vive and it has a lot better apps, so I wouldn't conclude about VR based on the Vive.

Don't they have the same resolution?

I actually want it for gaming but I'm pretending it's for work. ;)

Got my Vive last week. The other thing you don't realize is because of the lenses, the edges of your vision are blurry. So, you can't scan with your eyes like a regular monitor. I too wanted a virtual monitor and it's just not there yet.

That sounds like a dreadful future. Sitting for extended periods in virtual meetings with this headset on without any chance to refocus your eyes, drink something, doodle.

> That sounds like a dreadful future. Sitting for extended periods in virtual meetings with this headset on without any chance to refocus your eyes, drink something, doodle.

Ugh. I just imagined a room full of people sitting around a conference table with VR headsets on. Not a pretty thought indeed.

well that's quite sad. I was hoping a vr headset could be a valid substitute for multi-monitor setups, we just have to hope they get to that point. Smartphones kind of peaked around 4th gen, so we should wait a couple more gens of vr at least :p

I question how much resolution they have. How is the text rendering?

Their initial design was a fiber projector that oscillated in a spiral pattern with something like 512 rings (I can't remember the exact number). Along the the spiral you can have any resolution, only limited by how fast you can modulate the laser. Details along the spiral would be sharp, anything along the normal (spokes radiating inwards) would be only have 1024 distinct levels of detail.

But their later patents showed whole arrays of these fiber projectors. They are probably limited by the optics in-between the fiber and the eye and by the display controller bandwidth more than anything.

I had the same thoughts, but if its true!

But might was well stay hopeful for as long as possible.

Problem is that would require at least two huge leaps in technology one in display technology and one in compact computing power, probably more.

They've been doing this for how many years? And sinking how much money into content development? How about developing an actual product?

Isn't this what you dream of as a developer, the budget to do things right? Rather then pushing a demo into an actual commercial product?

Yes and no, you want to have a ton of time tinkering in a lab and the room to experiment and fail... but who would know Edison if he didn't eventually ship the lightbulb?

Also considering how litigious/secretive things are — it must suck to work on something for so long and be contractually obligated to avoid talking about it in public.

How long was the iphone in development?

About 5 years, but in working for Apple you worked for a company that still shipped other products.

Magic! Its in the name duh!

I assume the GTX 1080 lives in the cloud. We shall see I guess.

Can't. The bandwidth wouldn't be enough.

Bandwidth is not the problem, latency is. Bandwidth can be fixed by throwing more bandwidth at the problem.

Cloud gaming has not done all that well; it's right at the edge of feasibility due to latency. VR moves it back off the edge to "unquestionably infeasible". If VR gaming were to take off and become a dominant paradigm it would make cloud gaming even harder to justify.

Gaming and banking/trading are definitely the informative projects here.

Pseudo-real-time gaming is possible, but only for regional servers. Live global gaming is off the table permanently unless known physics changes; speed of light delays are larger than human reaction times. Right now, the best we can do is predictive techniques and continental servers, which have barely enabled MMO shooters.

VR gaming worsens the problem - 80ms delays in VR are nauseating and game-ruining. Global latency isn't a solvable problem via money or devotion.

> Bandwidth is not the problem, latency is. Bandwidth can be fixed by throwing more bandwidth at the problem.

There is a limited useful spectrum that's shared between everyone in the area. You can't just throw more bandwidth at the problem.

You sound like you think wireless is the only option...? All my consoles are wired. Anything that was doing cloud gaming would take an ethernet port unless the manufacturer is just an idiot. It might work with wireless but it would recommend wired.

I had to look it up, but it appears OnLive was shut down in 2015 and patents sold off to Sony.

I was completely skeptical at their claims with respect to latency. I never believed they could get the service to work, and for enough people to make it profitable. It appears quite a few people were able to enjoy the service at decent enough framerates. Which is surprising.

I used onlive for a year or so and really enjoyed it. There were some issues, but even the multiplayer FPSs were playable. Most of the time I was playing single player games that were less twitchy, and if you had done a blind test I probably wouldn't have been able to tell you if it was running locally or not.

I recently tried the PS4's remote play functionality. My friend in another city was able to play my bloodborne, which is a very timing-sensitive game, perfectly fine.

A few weeks ago my cousin ran games at 1080p on an EC2 GPU instance and used Steam's In-Home Streaming to play remotely and it worked perfectly fine.

I think the latency is now there, if you have a fast connection and live relatively close to the datacenter.

Latency is also critical for VR and AR

Latency is critical-path for VR not just in existing solutions, but fundamentally. Speed-of-light delays are sufficient to cripple remote-processing approaches to VR, so we won't see cloud solutions unless we're wrong about the fundamentals of relativity.

I wouldn't rule it out completely, though. There is some opportunity to mitigate latency, such as performing the heavy lifting remotely with local transformations on a simplified description of the scene to account for head movement and the like.

It would certainly be a much more complex solution that would still have inferior results compared to computing everything locally (just like cloud streaming now). Just that it's not "change the laws of physics" impossible to get to a workable implementation.

See also John Carmack's comments from 2013 on latency mitigation: https://web.archive.org/web/20140719053303/http://www.altdev...

critical but not impossible to fix. I can't find it but Microsoft published a video some time ago to realize cheap VR/AR headsets which do rendering in the cloud

A previous article (possibly from over a year ago) said that the computing power and battery come from a cell-phone-sized device that you keep in a pocket and attach to the glasses via fiber.

Uncompressed 8K over the public internet!

Why uncompressed? Recent mobile devices have hardware video decoding.

There are plenty of devices without HEVC hardware decode, though, which is basically a prerequisite to 4K. And software decoding is god-awful for anything besides South Park.

(It's not actually required, but the efficiency loss between H.264 and HEVC for 4K is pretty massive.)

Supposedly, because it's only rendering a small portion of your field of view at any given time, the compute resource required to power it is substantially less. I seem to also remember reading about it not using a standard pixel-filling kind of rendering, and using something else that requires less power instead.

Reply on news.ycombinator.com