Fool me, this is real good shit ! [the VR thread]

[venusian]

I bought an AMD R9 290 with 4GB RAM. Cost £215. Howls like a banshee when it's busy but I can't hear it with the headphone on :-)

[Soluzar]

I bought an AMD R9 290 with 4GB RAM. Cost £215. Howls like a banshee when it's busy but I can't hear it with the headphone on :-)

That was pretty much the best card to get for a while there, in terms of price / performance ratio. A significant step above mine, but then it's a few months newer I think.

I assume you'd be fine with most games whether VR or not on that card. There are some recent releases where no amount of power you throw at it will get the game to run perfectly though. I've heard reviewers who run 2x GTX 980 in SLI still can't get certain Ubisoft games to run nicely.

[KlaxonCow]

If it's "tethered" "it does not matter" at that point you can put all the grunt you want inside the PC.

Indeed, surely it makes more sense, engineering-wise, to try to work out how to make the thing entirely wireless - so it doesn't matter - than to try to work out how to strap an entire PC box to someone's head without it breaking their neck.


As for the cost of the system needed to run it, that's just the price of being an early adopter. You're essentially paying a premium for the privilege of being first, and that helps pay off the development costs.

It's, unfortunately, how it always works.

Remember those banking city "yuppies" with the ponytails in the '80s with their brick-sized mobile phones? They might have been tossers, but their early adoption of mobile phones created and fed the market, so that it could then move forward to create the wonderful array of cheap and cheerful phones that pretty much everyone now has in their pockets.

But it could never have feasibly and economically just jumped to the position we are in today from nothing whatsoever. R&D is expensive, things have been learnt along the way, companies need to be sure that there's a market to sell the bloody thing to when they've finished making it, etc.

It's a bastard, I'll grant you. I'd love a £30 supercomputer too. But these things take time to happen.

(The good news, though, is that they do often eventually happen. I mean, to someone dreaming that dream back in the '80s, the Raspberry Pi is exactly that £30 supercomputer, in relative terms, so it might well happen one day. But not today.)

In the meantime, you could always utilise Google's VR solution: https://cardboard.withgoogle.com/

[Dio]

2.5/3D die stacking?

How does that (dramatically) reduce either power consumption or cost? Stacking dies means still making all the dies before gluing them together.

[Absolute_Zero]

2.5/3D die stacking?

How does that (dramatically) reduce either power consumption or cost? Stacking dies means still making all the dies before gluing them together.

Smaller die [stacked] should have significantly better yields than the equivalent area on a single die.

In my naive mind:
- reduced voltage/frequencies
- increased bandwidth
- better yields

...and as the 3D technology matures, on to the other parts of the equation. e.g. stackable multi-CU die... stackable CPU cores for APU.

[Dio]

DRAM is a trivial case though because every chip is the same thing. Sure, you could turn every video card into multi-GPU...

[Soluzar]

DRAM is a trivial case though because every chip is the same thing. Sure, you could turn every video card into multi-GPU...

Aren't they already? I thought that's essentially what nVidia had been doing for a long time. Scalable GPU cores, of which the powerful cards get more. They also get more RAM, more pipelines and other stuff, too... but it's basically a scalable design. At least I think it is, and I won't mind at all if you want to educate me so I know better.

[gilesgoat]

DRAM is a trivial case though because every chip is the same thing. Sure, you could turn every video card into multi-GPU...

Yes but it would still be some "N" times the price I suppose.

[Dio]

Aren't they already? I thought that's essentially what nVidia had been doing for a long time. Scalable GPU cores, of which the powerful cards get more. They also get more RAM, more pipelines and other stuff, too... but it's basically a scalable design. At least I think it is, and I won't mind at all if you want to educate me so I know better.

Some percentage of the chip is made up of repeated compute units. It's well under 100%, even on the largest chips; things like display and physical interfacing is more than you might think, and there's elements that don't scale so well (like rasterisation) too. The buses that connect those things are gigantic as well, 1024-bits isn't that much interconnect really. So the ability to go wide at these points where the amount of data is very large is non-trivial.

Yes but it would still be some "N" times the price I suppose.

Chip manufacture isn't linear in cost, so more small dies do cost less than the same area in a large die. MGPU has its own efficiency concerns; dual-GPU and VR have obvious synchronicities though.

[gravy]

Doesn't that mean DRAM stacking is exactly the right thing to do for VR?

Is there any efficiencies to be had given the GPUs are basically operating on the same datasets with very similar access patterns to that memory?

[Dio]

Obviously I think DRAM stacking looks very interesting. After all, whose presentation is that slide above from? (I had to look up the presentation on the internet before I could even comment on it ).

Multi-viewport is interesting but non-trivial, and it would become more so if you are e.g. trying to share some large L3 between the GPUs. There are other techniques in the works too.

[gilesgoat]

Doesn't that mean DRAM stacking is exactly the right thing to do for VR?

Is there any efficiencies to be had given the GPUs are basically operating on the same datasets with very similar access patterns to that memory?

Probably, considered that VR rendering goes more or less like :

Code: Select all

BEGIN
GET_DATA_SET (A)
SET_VIEWPORT (LEFT_EYE)
SET_PROJ/WORLD/CAMERA/DISTORSION ( LEFT_EYE )
RENDER (A)
SET_VIEWPORT (RIGHT_EYE)
SET_PROJ/WORLD/CAMERA/DISTORSION ( RIGHT_EYE )
RENDER (A)
END

You could at the very least do two renderings in one time instead of two times.

Well given GPU parallelism and such hard to say "how it really happens" but sure that's a case where "two gpus should be better than one".

True.dat you may have one finishing a bit earlier/later than the other but still would be an almost 2x speed improvement ?

[Dio]

Merging the pixel shading is pretty hard if you think about it. As soon as you've got interpolation in the shader then you're dependent on the barycentrics, and those are going to be different for the two eyes. The final result may be very similar but it is slightly different (and I imagine if you tried to eliminate that difference you'd start seeing snapping or shimmering).

You can save bandwidth if you're synchronised, since you'd be hitting the same bits of textures, but I'm not sure how much compute you can save.

[gilesgoat]

Merging the pixel shading is pretty hard if you think about it. As soon as you've got interpolation in the shader then you're dependent on the barycentrics, and those are going to be different for the two eyes. The final result may be very similar but it is slightly different (and I imagine if you tried to eliminate that difference you'd start seeing snapping or shimmering).

You can save bandwidth if you're synchronised, since you'd be hitting the same bits of textures, but I'm not sure how much compute you can save.

I am just saying about having one GPU per eye rendering something that shares at least a common data set I absolutely am NOT saying that the two GPUs will be rendering 'the same thing' but could be an improvement to have one per eye ?

For this kind of stuff you should really get in touch with people like Carmack or such .. I know that when into VR "there are lot more issues than standard rendering" .. the big thing would be to make some things happen "while display is going on" or while rendering is still going, for the little I understood/know about this some stuff maybe "works better in some other GPUs" where you can "mess up" with some stuff cause the "macrotile" arrangment ( i.e. I mean before another macrotile gets rasterized/rendered something can be done ).

Fundamentally if I understood correctly is that "the big issue" is that 60 hz .. or 75 hz .. or even 90 hz .. is "far from perfection" unfortunately you can move you head and cause accelerations and stuff much faster than this, some say 300 Hz is what you'd need you can't really ( should not ) wait "a full frame" ( at 60 or 90 ) being displayed before you can do anything else with the sensor data, you can continue to be moving your head quite a bit during the time the display is just showing its scanlines.

Another really big issue is the quality of the display itself, response time have to be "stupidly fast" or "anything turns into blur".

But the display itself "can be done" the rendering is the problem and the possibility to "correct while display is showing" ( which basically means "a partial re-rendering of the stuff while scanlines are going" ) is the big problem.

And then of course resolution and colours and such.

I think you'll have a fuck load of work into "make things even better/more powerful" for quite a number of years to come, which I think should be interesting for you because I had the impression that the development/etc. into graphics cards for PCs was a big "stagnating"

[Dio]

The stagnation thing is a technology one. The pace we can innovate at has slowed a bit as the process nodes and frequency jumps aren't cascading along at the rate they were. Moore's Law is long gone, I fear, although we're still managing to generate steady growth.

For this kind of stuff you should really get in touch with people like Carmack or such ..

We are well plugged in up and down the whole software stack.

On a similar topic I had a pint with Callum last week and he was very impressed with TxK and very surprised I did a part of the soundtrack .