Why neuromorphic computing?
We've discussed the many differences between the brain's computation and how a traditional CPU works several times at ExtremeTech, but the big-picture take-away is this: While modern computer processors are multiple orders of magnitude better than the human brain at certain tasks, they're far less efficient then the brain at others.
Building processors that simulate how neurons fire, connect, and share information with each other is a formidable undertaking, but Qualcomm has been working on its Zeroth platform (also known as an NPU) for the past several years. The company gave a talk on the NPU architecture last year (a block-level diagram of the proposed architecture is shown below).
The chip integrates a scalable number of neurons, but it's difficult to comment on the design architecture shown here -- or to even say how it materially differs from IBM's TrueNorth, the one-million neuron processor with 4096 cores and 5.4 billion transistors. The one thing we can definitively say is that any NPU that ships inside a mobile phone will, by necessity, be much smaller than IBM's enormous core. Qualcomm could theoretically be developing a server backend to augment smartphone capabilities, but if it is, it hasn't said so yet.
What Qualcomm is saying is that the Zeroth platform will have the ability to learn from a users' actions and can transfer that learned knowledge across different devices, even when the user upgrades a smartphone. The company is pitching Zeroth a bit like a smart administrator that knows your preferences before you do, and can leverage its own measurements and capabilities to take better photos; switch intelligently between WiFi, Bluetooth, and cellular data depending on which signals are the strongest; and use "advanced behavioral analysis" to detect and protect against certain types of malware. Zeroth is also supposedly capable of recognizing gestures, expressions, and faces, and intelligently sensing its own surroundings.
Separating hype and fact
It's difficult to tease out where the practical capability of the Zeroth platform starts and the hype ends. For one thing, this kind of intelligent monitoring and practical decision-making capability is precisely what could make the Internet of Things practically useful as opposed to kludged and annoying. The Holy Grail of smart sensors and the so-called "Smart Home" are sensors that are actually smart. But even well-received devices like the Nest thermostat merely scratch the surface of what hardware might one day accomplish.
This slide is from 2014 -- the new integrated announcement implies availability has improved
The problem with this kind of hype, however, is that programming neuromorphic chips is nothing like writing code for any other kind of application or sensor. The degree of heavy lifting that needs to happen before chips like the Zeroth are ready for seamless integration may be substantial.
Still, it's interesting to see Qualcomm taking a shot at this kind of integrated processor. While I don't expect the Zeroth platform to reinvent computing, the only way to build products for a nifty processor that attempts to mimic certain organizational principles of the human brain is to build it and shove it out for folks to play with. From that perspective, the Zeroth could be a huge hit -- even if it'll take a few years before the chip's concepts are put to practical use.
