Ok so decided to give this a full read and thought that it might be worth going into a bit of detail.
The recent presentation provided a bit of a clue and certainly spiking neurons is the future. Moving away from convolutional networks to spiking neurons seems to be a consensus moving forward.
So NYU labs dwelled into developing convolutional networks and did advance a little to arrive at few Proof of concepts such as obstacle detection. I would hazard a guess that they did much more than a POC as there are quite a few research papers by the same team around object recognition and segmentation.
Their primary argument has always been that spiking neurons creates a binary output. So a returned result would generally be 0 or 1. Let's say you were trying to calculate whether a given object is a dog.Hence in order to get precision like whether a perceived object is a dog , they require multiple cycles of the processor to exactly compute whether the perceived object is a dog because the output's are binary (0 or 1) and we require precision.
So IBM has a been the header in this area with announcing some progress in their
ambitious project. Clearly brainchip are not the first movers in this field. The problem with IBM has though been that they have been able to theoretically do a lot of things but when it comes to a usable implementation they have always been in doldrums. Just recently IBM dropped yet again in stock market in relation to their falling revenues.
Anyways , IBM's Truenorth has 4096 cores and can perform 46 billion operations per second while a normal i7 or a xeon processor on the market performs close to 1 bil or less. The other advantage is that the processor is scalable. It performs the operations in parallel as opposed to simultaneous processing of our generation.
So when the Truenorth processor was challenged it delivered exceptional output.
Now what's primarily stopping brainchip are a few impending questions that when answered will see Brainchip generating substantial interest
FPGA usually consume more power. Peter's design for POC uses FPGA.So what is the actual power consumption per a processing cycle or 1 billion operations ? Given the current access and scarse information I am presuming that they are still processing information in each time step. Otherwise they would clearly go about announcing this as their USP.
64 chip ? How ? When ? Other Designs specs ?
BDK they will build and deliver in 2 years ? Whats the implementation of their BDK involve ? Is this going to be open source ?
What are the experimenters kit ? will these be open by design ?
What sort of encoding does the processor use ? Is this going to be boolean encoding ?
Neural spiking would obviously have the core advantage of power saving but obviously Brainchip will have the first mover advantage in this market apart from obviously IBM and a couple of others.
A lot of iteration and how motivated is the team will dictate the success of this project. Either way I will predict that in 5 years this company will either be acquired by a giant like GOOG , FB or the likes or will be establishing a stream of revenue that will make it too big to gobble.
It's a good buy opportunity and worth every penny at 32 cents.
(20min delay)