Latency worries me, though. Bandwidth has been improving a lot faster than latency for a while now. For always-on augmented reality, I think that we’re going to need some seriously more power-efficient computing so we can do latency-limited tasks locally. (Also, communication takes energy too—often more than computation.)
Good news on that, by the way: modern embedded computer architecture and manufacturing techniques are going in the right direction for this. 3D integration will allow shorter wires, making all digital logic much more power efficient. Network-on-chip architectures will make it easier to incorporate special-purpose hardware for image recognition and such. And if you stick the memory right on top of your processor, that goes a long way to speeding it up and cutting down on energy used per operation. If you want to get even more radical, you could try something like bit-serial asynchronous processors (PDF) or something even stranger.
Latency worries me, though. Bandwidth has been improving a lot faster than latency for a while now. For always-on augmented reality, I think that we’re going to need some seriously more power-efficient computing so we can do latency-limited tasks locally. (Also, communication takes energy too—often more than computation.)
Good news on that, by the way: modern embedded computer architecture and manufacturing techniques are going in the right direction for this. 3D integration will allow shorter wires, making all digital logic much more power efficient. Network-on-chip architectures will make it easier to incorporate special-purpose hardware for image recognition and such. And if you stick the memory right on top of your processor, that goes a long way to speeding it up and cutting down on energy used per operation. If you want to get even more radical, you could try something like bit-serial asynchronous processors (PDF) or something even stranger.