Microsoft HoloLens: HPU Architecture, Detailed

[cangelini]

Essentially, the HPU is designed as a companion to the SoC, receiving all of the myriad sensor data, packaging that data, and feeding it to the SoC. That's what it does, but how does it work? Let's take a deeper look at the HPU's architecture.

Microsoft HoloLens: HPU Architecture, Detailed : Read more

 

[wifiburger]

yes finally some information on their product, we all know people have to know what's inside this turd ! that's what sells it and makes Microsoft money NOT !

 

[bit_user]

Thanks for the update.

Initially, Microsoft didn’t want any one compute node utilized more than 50%, so there should be plenty of headroom built into the HPU for more taxing algorithms, thereby future-proofing the HoloLens as much as possible.

The company also specified exacting latency and duty cycle requirements, intended to reasonably assure the best experience possible.

In order to satisfy realtime latency requirements, they need to leave some headroom. So, they can't just double the workload and still hit their latency targets.

To me, the 1 GB of for the Atom cores seems like the weakest spec. This thing has to run Windows 10, their graphics stack, their AR stack, and the AR app, itself. For all that, 1 GB seems tiny.

 

[alextheblue]

Bit - the HPU does a lot of the heavy lifting and that has it's own RAM. Also although it's running Win10, it's not configured the same. The memory footprint is greatly reduced, anything not needed isn't running.

 

[bit_user]

Bit - the HPU does a lot of the heavy lifting and that has it's own RAM. Also although it's running Win10, it's not configured the same. The memory footprint is greatly reduced, anything not needed isn't running.

Even if you strip down Win 10 to the minimum, once you add back in the stuff needed for AR + apps, it seems like 1 GB ought to be a squeeze.

I'm confident they'll add more RAM, in the next version. Maybe that'll even be something I can afford!

Hopefully, it'll also use Broxton, based on Intel's Goldmont core, which is allegedly 50% faster per clock. That also has Gen9 HD Graphics (up to 18 EUs, I think).