Comment by InTheArena | Hacker News Reader

Comment by InTheArena | Hacker News Reader

So the new Apple iPhone X has a front sensor that can generate depth maps, and a rear-sensor that does even more sophisticated depth mapping. They also have a CPU that is as powerful as anything you get in Desktop PCs.

So tell me, what's to keep someone from using the depth information on the front camera, using a Google Dream VR setup (but instead of covering the back of the iphone, use that sensor for the world around them) and then projecting poor man's AR with today's world. If the answer is wrong sensors, they can fix that next year, and do it at a size and a scale that no one else can touch.

You just took a hardware problem, and made it a software problem.

The fact that Peter Jackson has evidently jumped ship to ARKit is a pretty big sign that Apple is going to get there first.

Apple's focus on ML on the phone directly, not requiring a server at the RTT latency is perfect for this application. They have hardware no one can touch, and do full stack miniaturization like no one else. Plus they also attract "lifestyle" buyers.

The rear sensor does not do "even more sophisticated depth mapping". TrueDepth sensor is only on the front. The CPU is not as powerful as anything you get in Desktop PCs. There's an order of magnitude power difference between a mobile SoC and a desktop CPU/GPU combination. If you have doubts about that, try running full Photoshop on a mobile SoC via Windows 10 on ARM. Not sure what "Dream VR" is. Do you mean Daydream?

Nothing is keeping someone.

Probably the depth sensor is not high enough resolution. The camera is not high enough resolution. The phone battery will be dead in 45 minutes. I imagine the high end AR market would be for things like surgery where extreme precision matters. For instance hook a high precision AR device up to a C-arm and the way angioplasty is carried out would be completely changed forever.

While AR would be a cool use-case for angioplasty the difference to current methods (https://en.wikipedia.org/wiki/Interventional_radiology) probably isn't as big as you think.

I guess I meant data-wise. The image of the catheter and blood vessels would be superimposed in the AR device and the doctor would look at the patient instead of a monitor.

By putting AI on the Phone, they invert the data center and communication problem. The mainframe vs smart client cycle continues. A lower powered chip on the client would need a more powerful chip in the datacenter to make up for data transit time. Individual users pay for power, manage heat dissipation and upgrades. This actually saves Apple billions in data center costs while making them money.

And as a nice add-on, they get to brag about being privacy friendly since nothing leaves the device.

I'm becoming quite convinced that long-term this might be Apple's biggest selling-points (and they realize this).

In precisely the areas where Apple has or might end up having difficulty competing with the likes of, say, Google and their massive datasets, the focus on privacy and security might very well be enough for consumers to go for the 'inferior' product.

For example, I've considered switching away from Apple for my new phone, but the security of the iPhone (or so I've read multiple times here on HN) is enough to keep from even considering a switch to an Android device even if everything else were better on Android.

The same goes for my photos. I'd love the features Google Photos offers, but I just don't feel comfortable uploading them all to their servers. While Apple's implementation of these features might be inferior, at least I can keep things on my (secure) devices.

> They also have a CPU that is as powerful as anything you get in Desktop PCs.

Is this actually true?

Yes it is, at least when using Geekbench as a measure. The A11 beats some of the Intel processors in the 2017 MacBook Pro models.

https://www.macrumors.com/2017/09/13/a11-bionic-chip-geekben...

Well, that makes the CPU as powerful as some processors in mobility focused laptops, not as powerful as "anything you get in desktop PCs".

I'm writing this from my linux desktop. It has 11k/3.8k for the geekbench numbers. This is pretty much there.

No.

Geekbenchmark is a bad test. It does a series of short burst tests spaced out over time. These avoid harsh thermal throttling on mobile devices, yet on larger desktops they don't register as a heavy workload so the maximum clock doesn't come into play. Bigger chips will stay down clocked.

Effectively what Geekbenchmark's result ends benchmarking sleep state transition speed.

Desktop/Laptops don't spin up/down as fast as cellphones so this heavily favors them. Just doing a continuous long test would heavily favor desktops/laptops as they have better designed heatsinks.

---

iPhones do have great processors, just fair mobile/desktop test are very hard. They are different chips, that require different power, change clock states in different ways, and have different thermal constraints.

The core workloads in most scenarios are very different, so fair for one isn't fair for another.

I might be talking from ignorance here, but if you put the iPhone chips in a different environment that included the same active cooling that desktop CPUs enjoy, would this still be true?

Most arm servers perform incredibly poorly per core and are significantly less power efficient.

Unlikely.

Modern X64 cores are super smart. Not just on _easy_ things, like predicting branches, and how data flows though instructions. But with caching, memory access, inter-core communication, cache snooping, and cache synchronization/invalidation.

The real issue for modern system is RAM access. Everything you CPU does between hitting RAM may as well be instant how expensive a RAM hit is.

X64 has a large cache hierarchy to make this rarer. A11/A10X/A10 less so.

Transistor count is a limited metric. That said, the A11 supposedly has 4.3 billion transistors, which is more than IBM's 12-core POWER8 chip's count. Apples to oranges, since the A11 has far more application-specific circuitry and much lower power consumption & clock speed.

https://en.wikipedia.org/wiki/Transistor_count

Ps: should be noted that the embedded GPU in the A11 contributes many transistors.

Tend to agree. The big problem no-one has solved is the glasses/immersive aspect.

Hololens is close, but the kit is waaaay to bulky still (and low FoV, not bad tho). In a few years tho... some kind of ARKit accessory will absolutely kill in this space.

Yup. Apple's depth sensor + ML can be also used as pretty powerful 3D hand gesture mapper. Better than Leap Motion I would guess. Also, switching to OLED allows low persistence. Their new custom GPU and custom video encoder on iPhone 8/X might even allow them to do ultra-low-latency AR pipeline. It's coming it's for sure.

getting there first is not the same as winning :)

Specially when you have players around with unlimited cash.

Reply on news.ycombinator.com