This site may earn affiliate commissions from the links on this page. Terms of employ.

MIT-NeuralNetwork

Neural network processing and AI workloads are both hot topics these days, driving multiple companies to announce their own custom silicon designs or to plug their ain hardware as a top-end solution for these workloads. Only ane problem with neural networks is that they tend to exist extremely power intensive, and not necessarily suited to mobile devices or the kind of low-power "smart" speakers that have recently get so popular.

MIT is claiming to take developed a neural network processor that addresses these bug, with an overall power reduction of upwardly to 95 percentage. If true, this could change the game for these kinds of applications. Instead of being forced to rely on cloud connectivity to drive AI (and using power to keep the modem agile), SoCs could incorporate these processors and perform local calculations.

"The general processor model is that there is a retentiveness in some part of the flake, and at that place is a processor in another part of the chip, and you move the data dorsum and forth between them when you practise these computations," said Avishek Biswas, an MIT graduate student in electrical engineering and informatics, who led the new chip's development:

Since these car-learning algorithms demand and then many computations, this transferring back and forth of information is the dominant portion of the energy consumption. Just the computation these algorithms practise can be simplified to one specific operation, called the dot production. Our arroyo was, can nosotros implement this dot-product functionality inside the memory and then that you don't demand to transfer this information back and forth?

A typical neural network is organized into layers. Each node connects to other nodes above and beneath it, and each connection betwixt nodes has its own weight. Weight, in this context, refers to how much of an bear upon computations performed in 1 node will have on the calculations performed in the nodes information technology connects to. Nodes receiving input from multiple nodes higher up it multiply the inputs they receive by the weight of each input. The result is called the dot production. If the dot product is above a certain threshold, it gets sent along to nodes farther down the concatenation. But this process is extremely memory intensive, with each dot product calculation requiring memory accesses to retrieve the weighted values. Those values and then have to exist stored, and each input to a node has to be independently calculated.

What MIT has washed is create a flake that more closely mimics the human brain. Input values are converted to electrical voltages, and so multiplied by appropriate weights. Only the combined voltages are converted back into digital representation and stored for processing. The prototype chip tin can calculate sixteen dot products simultaneously. Past storing all of its weights equally either one or -1, the system can be implemented as a elementary fix of switches, while only losing ii-3 pct of accuracy compared with the vastly more than expensive neural nets.

Not bad for an approach that can reduce ability consumption upwardly to 95 percent. And information technology's a promising concept for a time to come in which the benefits of the cloud and AI aren't limited to those with robust internet service in a mobile device or at domicile.

Meridian image credit: Chelsea Turner/MIT