Intel discusses neural network inside Quark MCU
The information comes in a ‘white paper’ (see Intel: Why an x86 MCU is right for IoT).
Intel introduced the QuarkSE microcontroller – with brief details of its pattern matching technology – on the ‘Curie’ circular button-sized printed circuit board at the Consumer Electronics Show in January 2015 (see Startup’s tech is Intel’s Quark neural network).
Bill Giovino, author of the recent white paper, states that the pattern-matching recognition engine is a network composed of 128 arithmetic units or neurons designed to perform two types of pattern recognition; the k-nearest neighbour (KNN) recognition and the radial basis function. Up to 32,768 patterns can be programmed and engine returns a positive match, uncertain, or negative match within a fixed time, he states.
Giovino’s suggestion is that the network is used as part of a wake-up system so that a sensor subsystem can pass a series of vectors to the PME, which matches it against a stored dataset. If there is hit this can be used to send a wake-up event to the x86 core. This can in turn then decide whether to process information locally, perhaps using a 32-bit ARC DSP that is present on the MCU die or to send information on to a sensor hub.
The pattern-matching engine is based on technology developed by General Vision Inc. (Petaluma, Calif.), developer of the CMIK chip in 130nm process technology with 1024 neurons and 256 bytes of memory per neuron. The Quark SE has 128 neurons and 128bytes of memory per neuron and is implemented in 22nm process technology, according to General Vision.
Related links and articles:
Intel: Why an x86 MCU is right for IoT
News articles:
Startup’s tech is Intel’s Quark neural network
NeuroMem put on hold as LabView calls
Startup wants to be the ARM of neuromorphic cores
BrainChip provides details of neural network architecture
If you enjoyed this article, you will like the following ones: don't miss them by subscribing to :
