HoloLens 2 can learn!
Microsoft announced that the second generation of the HoloLens' Holographic Processing Unit (HPU) will contain a deep learning accelerator. When Microsoft first unveiled the HoloLens, it said that it comes with a special kind of processor, called an HPU, that can accelerate the kind of "holographic" content displayed by the HMD. The HPU is primarily responsible for processing the information coming from all the on-board sensors, including a custom time-of-flight depth sensor, head-tracking cameras, the inertial measurement unit (IMU), and the infrared camera.
The first generation HPU contained 24 digital signal processors (DSPs), an Atom processor, 1GB of DDR3 RAM, and 8MB of SRAM cache. The chip can achieve one teraflop per second for under 10W of power, with 40% of that power going to the Atom CPU. The first HPU was built on a 28nm planar process, and if the next-generation HPU will be built on a 14/16nm or smaller FinFET process, the increase in performance could be significant. However, Microsoft has not yet revealed what process node will be used for the next-generation HPU.
What we do know so far about the second-gen HPU is that it will incorporate an accelerator for deep neural networks (DNNs). The deep learning accelerator is designed to work offline and use the HoloLens' battery, which means it should be quite efficient, while still providing significant benefits to Microsoft's machine learning code.
(Score: 0) by Anonymous Coward on Tuesday July 25 2017, @01:34PM (4 children)
I'm not sure what they plan to do with this... it will be too slow to train unless it is very small. I guess a lot of these techniques were devised on even slower devices though (but then you would wait weeks for a result).
(Score: 2) by takyon on Tuesday July 25 2017, @02:10PM
That is the amount for the first generation device, not the second (amount unknown). HoloLens 1 actually has 2 GB of RAM total. 1 GB is dedicated to the "holographic processing unit".
[SIG] 10/28/2017: Soylent Upgrade v14 [soylentnews.org]
(Score: 2) by Nerdfest on Tuesday July 25 2017, @03:38PM (1 child)
What they plan to do? Based on recent experience .... "telemetry".
(Score: 2) by Gaaark on Tuesday July 25 2017, @04:34PM
Based on my 'ancient' experience.... "BSO telemetry".
--- Please remind me if I haven't been civil to you: I'm channeling MDC. ---Gaaark 2.0 ---
(Score: 0) by Anonymous Coward on Tuesday July 25 2017, @04:44PM
Maybe it's good for recognizing squirrels? (yesterday's article about using a Pi for scene recognition)
(Score: 2) by ledow on Tuesday July 25 2017, @04:45PM (3 children)
So they stuck "AI" in a VR headset.
I mean, yeah, it's a buzzword search-term grabber.
But really? What's the point? What does it "learn"? (P.S. it doesn't, even the "deep learning" / "machine learning" stuff is still just snake-oil, especially on this scale).
(Score: 3, Interesting) by JNCF on Tuesday July 25 2017, @06:21PM (2 children)
My first thought was predictive rendering. We experience lag even without augmented reality. Our brains try to adjust for this, sometimes with mixed results. [wikipedia.org] Some modern online games also try to predict how their worlds should be rendered by anticipating the actions of distant players (not just assuming a continuation of the last known input), rendering the world as if those actions happened, and then rolling back the changes if those predictions were wrong. See GGPO, for example. [wikipedia.org] Given that even a teeny bit of lag would be annoying in augmented reality, I could see wanting to predict eye and head movement based on previous experiences with a given player. If 98%* of the time I glance left and then right in a given time interval I then proceed to glance left again, it might be helpful to just assume I'm going to do that and roll back the changes if necessary. While predictive rendering introduces some glitches, it can be tuned to prevent more errors than it introduces.
*Number pulled directly from ass.
(Score: 0) by Anonymous Coward on Tuesday July 25 2017, @06:33PM (1 child)
> While predictive rendering introduces some glitches, it can be tuned to prevent more errors than it introduces.
In talking to high end simulator suppliers (the multi million $$ simulators used in F1 and other professional racing), one consistent comment is, "Never give a false cue." I believe this was originally relating to motion cues (moving simulator base), but probably applies to visual and audio cues as well. Better to do nothing (leaving a hole for the brain to fill in?) than to explode (mentally) the simulation with something that is wrong.
(Score: 2) by JNCF on Tuesday July 25 2017, @06:47PM
(Score: 2) by requerdanos on Wednesday July 26 2017, @01:50AM (1 child)
One Tera Floating Point Operations Per Second per second?
I know that idiotic nonsense like "PIN Number" and "ATM Machine" redundancize the world already, but flops per second is not a welcome addition.
(Score: 0) by Anonymous Coward on Wednesday July 26 2017, @06:13AM
It's the Amazon AWS growth rate.