Neuromorphic processor

from Wikipedia, the free encyclopedia

A neuromorphic processor ( English : neuromorphic processor unit ; NPU ) - also known as a neurosynaptic processor (English: neurosynaptic processor unit ), short neural processor ( borrowed from the English neural processing unit ) or AI processor - is a processor whose architecture is based on Neuromorphing based.

application

Neuromorphic processors are Turing-complete and therefore universally programmable. However, they are particularly suitable for tasks in pattern recognition and analysis .

Typical algorithms are used to build convolutional neural networks (CNNs), liquid state machines (LSMs), restricted boltzmann machines (RBMs) or hidden Markov models (HMMs), as well as looming detection, the estimation of spectral power densities and the temporal Pattern matching .

advantages

In contrast to the Von Neumann and Harvard architecture  , neuromorphic processors are not subject to the Von Neumann bottleneck . They are particularly scalable.

Since neuromorphic processors are fault tolerant, neuromorphic processors can be manufactured with a higher error rate. This enables z. B. to manufacture larger processors with lower reject rates.

Another advantage is the lower energy consumption, since the individual neurons are event-driven and therefore only need energy occasionally. They are also suitable for controlling neuromimetic chips that are used in brain-computer interfaces .

disadvantage

Neuromorphic processors have the disadvantage compared to Von Neumann processors that the representation of numbers and exact mathematical calculations are only possible with great effort. However, this can be compensated for because neuromorphic processors can work together with conventional processors.

In addition, programming languages, compilers and operating systems optimized for the hardware must first be developed.

Implementations

Single receipts

  1. ↑ AI processors: Why the new chips are the future - PC Magazin , February 21, 2018
  2. ↑ Steve K. Esser, Alexander Andreopoulos, Rathinakumar Appuswamy, Pallab Datta, Davis Barch, Arnon Amir, John Arthur, Andrew Cassidy, Myron Flickner, Paul Merolla, Shyamal Chandra, Nicola Basilico, Stefano Carpin, Tom Zimmerman, Frank Zee, Rodrigo Alvarez -Icaza, Jeffrey A. Kusnitz, Theodore M. Wong, William P. Risk, Emmett McQuinn, Tapan K. Nayak, Raghavendra Singh, Dharmendra S. Modha: Cognitive Computing Systems: Algorithms and Applications for Networks of Neurosynaptic Cores. (PDF) IBM Research ( Almaden ), IBM Research ( India ), University of California, Merced , accessed on August 8, 2014 .
  3. ↑ Brain Power. IBM Research, accessed August 8, 2014 .