Hacker News Reader

Comment by dsr_ | Hacker News Reader

Well... if you have 20/20 vision, you can resolve about one arc-minute. (That's at your fovea. Everywhere else, not so great. But we will probably find it easier to render everything than track your eyei saccades and re-render fast enough to fool them.) Your field of vision is about 180 degrees wide (not really) and about 135 degrees vertically (full eye movement).

So: 60 arc-minutes to a degree, 180 wide: 10800 pixels.

60 * 135 wide: 8100 pixels.

An 8192x4096 is not there, but a 16384x8192 is.

Now double that so you can get stereo vision (though really there's a lot of overlap that's going to be hard to arrange), update it 100 times a second in 48 bit color... 420 billion bits per second. Double that if you believe that we need 200Hz updates.

View on HN

Yes, but shouldn't we eventually be eye-tracking anyways for depth-of-field effects? If you're eye-tracking already, you can leverage that for resolution optimization.

You still need the display to support the resolution even if you can optimize rendering costs.

To allow the eyes to actually refocus based on distance you want light field displays in your VR headsets

How many years away are we from being able to push this kind of throughput through a GPU?

Oh boy, napkin math! (not sarcastic at all - I think this kind of ballpark estimate is actually super fun)

A Titan XP/1080 Ti is finally just about enough to push 4K/60Hz with a single card. 16K would be 16 times as many pixels per frame, and 144 Hz is more than twice the refresh rate.

To put that another way, we haven't even started to think about the connector standard that you could gang together to push that many pixels, let alone having a GPU that could actually push them. At this point you are pretty much talking about some kind of lossy compression being involved ("visually lossless" sure is a great euphemism). With lossy compression you might be able to get away with ganging together a couple of whatever DisplayPort 1.5 or 1.6 end up being.

In fact with the end of Dennard scaling there's some fundamental problems with how you would even use a GPU like that as a consumer. Your average (US) household circuit is 120V/15A peak, and can be run at 80% load continuous (12A), which works out to 1440 watts at the wall. With an 80% efficient PSU that works out to 1152 watts inside the case.

A Titan XP/1080 Ti pulls 270 watts at stock clocks [0]. So hypothetically even if you stacked four of them on an interposer you're now pulling 1080 watts inside the case, which is almost your entire circuit capacity. So by a naive calculation (16 times the pixels twice as fast = 32x divided by four GPU dies) we need at least an 8x improvement in overall efficiency before this is viable.

(Four big GP102-sized chips on an interposer is technically possible right now - they only need to be on the interposer where you need interconnect bumps - they can hang off onto a support substrate, and you can have shared memory controllers/HBM2 stacks/etc on the interposer that make this appear to be one big GPU chip instead of SLI/Crossfire. Since the interposer is actually a chip all on its own, some of these auxiliary functions can actually be built into the interposer itself (the "active interposer" concept), with the biggest obstacle being getting the heat out since the interposer has other chips stacked on top of it...)

220V users obviously have things a little easier here (~twice the capacity per circuit). This cuts the overall efficiency improvement necessary down to 4x.

There will also be some efficiency improvements on the software side. For starters we can cut quality a bit (that's 4K/60fps ultra), foveate rendering, and other software magic. Speeds don't scale perfectly negatively with increased resolution so you will get some savings there. Assume some speedup from DX12/Vulkan too.

Oh and since this is 16K/144 Hz per eye I guess there's an implicit assumption here that we can make NVIDIA-style Multi Viewport Rendering work at near-100% efficiency, otherwise that's another factor of (up to) 2x that needs to be accounted for.

So let's say that we need a 4x improvement in overall hardware/software efficiency. Let's say, something like a doubling in GPU throughput-per-watt and to double software efficiency before an absolutely state-of-the-art rig could even feasibly do this task on its own dedicated 220V/15A circuit. And there's a few fairly optimistic assumptions built into that 4x number.

Take your best shot at how long it will take to quadruple efficiency in the post-Dennard era. Let's say 20-30 years, unless there's a massive breakthrough in materials science or optical computing or something.

[0] https://www.techpowerup.com/reviews/NVIDIA/GeForce_GTX_1080_...

> 220V users obviously have things a little easier here (~twice the capacity per circuit). This cuts the overall speedup necessary down to 4x.

Actually, in Europe, you get 220V/25A from normal wall sockets, or, for stuff like washing machines or stoves, you get up to 400V/63A.

With that – roughly 3.5kW from a normal socket, or 25kW if you use a socket for a large appliance, compared to the 1.4kW available in the US – you can easily power a fuckton more.

Apparently the 400V/63A one also exists in 125A and 200A versions though I guess those are not very likely to show up in households. There's also a 690V edition. It's quite a shock.

Normally "large appliance sockets", as he put it, are 400 V / 16 A CEE, though, 32 A CEE is pretty rare in households and more a commercial thing, although you can install it, since total capacity is at least 40 A -- but usually 63 A or 100 A.

690 V is industrial (this the logical progression dictated by star-delta connection of motors), and not used in residential contexts. 125 A CEE is the largest standard size. Beyond that other connectors are used. Industrial use also sees other connectors / kinds of connections for higher voltages (up to medium voltage [couple kV]).

Ugh, you have no idea how annoying that is. I'd like to set up a little MPI/CUDA cluster for personal/friend use but four GPUs (1060/1070) at full tilt works out to 500-600W right there, which means a standard circuit only runs two boxes maximum.

As far as I can see it's either colo or run a 220V circuit, which is a bit of a non-starter for a hobby project.

Heck, you can even get uncomfortably close to maxing out a circuit with a single box. SLI 1080 Tis = 540W, plus 250W for an OC'd Intel HEDT processor is 800 watts inside the box, or 1000W at the wall (8.25A of 12A continuous). Hope your wife doesn't start the hairdryer while you're gaming...

How are you even using vacuum cleaners? Many vacuum cleaners here are 1800 or 2400W, kettles are 3500W, and hair dryers usually above 1800W, too.

This all works because you have between one and three 400V/63A triphase circuits per housing unit, but if your US circuits are this tiny...

Guess why it's 3500W max? Because Schuko is specced for 16A only (and 10A continuously). Where did you get that 25A number from?

The 25A is, in the house of my parents (built '05), what each breaker and line support.

As a line might be hardwired to multiple sockets, or to an appliance, or might be connected via a CEE plug instead of Schuko, it might support more than 16A.

If you build a large-scale GPU Cluster, you won’t connect them all to a single socket, but each to a separate socket, although they might be on the same line. That’s where the 25A can be useful.

I'm even more impressed with my eye and brain circuitry now.

The thing to remember about eyes is that they don't work in terms of frames and pixels.

This is really the fundamental problem with saying "X ppi is enough for the eye" or "you can't see faster than X Hz". This is a digital approach to an analog system. Your brain is tuned to identify resolution in the center of your vision and movement in the periphery, and has all kinds of "special case" circuitry to react quickly when necessary.

Classic example, fighter pilots are capable of recognizing silhouettes flashed on a screen at some ungodly rates. But if you were, say, playing a game, 100 Hz is very smooth already (even CounterStrike), and you might not be able to visually distinguish that from 144 Hz.

It's like a psychovisual model with a temporal aspect incorporated.

Thanks to the sampling theorem, if there is a frequency limit in a continuous signal, then you merely have to sample at twice that frequency.

The eye clearly has frequency limitations, so we just need to find the sampling strategy which accurately codes all perceptible frequencies.

It's going to be more than twice--Nyquist-Shannon is for 1D; 2D images would have a lower resolution on the diagonals, so you'd want a minimum of twice the frequency on the diagonals.

In an ideal world, we would also be using hexagonal grids rather than square, which gives a resolution increase for the same area and a given pixel size, and also reduces off-axis aliasing.

Very astute, thank you for bringing this to my attention. So it'd be about 2 and 58/70 times the maximum frequency, assuming two-dimensional sampling.

As for hexagonal grids, there are a number of complications introduced by that, especially when it comes to using general algorithms based on regular 2D euclidean space.

With a sampling rate of 2 on the diagonal, you'd have a rate of 2.828 on the axes, with a square grid (2 * sqrt(2)). Unless I'm missing something.

Agreed about the complications of the hexagonal grid. We make implicit assumptions everywhere. Though I'm not sure they are insurmountable for some purposes; with graphics APIs like OpenGL, texture coordinates and samplers could be using a hexagonal grid relatively transparently. That would allow use of hexagonal displays even if the texture data is square.

  The eye clearly has frequency limitations

No...the eye doesn't work that way. It is not a machine. It is organic and analog. It doesn't have some frequency limit. In any case, you're abusing the Shannon-Nyquist theorem. It absolutely does not imply that if some device is sampling at x hz then you can sample at 2x hz and produce a sufficient reproduction for the 1x hz sample. This is a common misconception.

Reply on news.ycombinator.com