MATHEMATICAL PROBLEM SOLVING CIRCUIT INCLUDING RESISTIVE ELEMENTS

Data di pubblicazione

02-12-2019

Codice

DEIB.17.068.A

Stato

Licenziato

Data di priorità

01-01-1970

Fase

Italian and PCT

Titolare

Politecnico di Milano

Dipartimento

Department of Electronics, Information and Bioengineering

Autori

Ielmini Daniele, Pedretti Giacomo, Sun Zhong

Descrizione

LOOPUS is a breakthrough in computing systems, enabling scalable, zero-latency, energy-efficient solution to machine learning and big data analytics via analogue in-memory accelerators. In particular it develops analogue accelerators for algebraic computing, delivering your machine learning and big data analytics in just one click.

Loopus is an innovative hardware accelerator for machine learning (ML) and processing of big data. It is developed a novel electronic circuit to train ML algorithms, including linear/logistic regression, neural network and page ranking in one step. The circuit is based on non-volatile analogue memories and feedback systems (thus the name Loopus).

It been developed the concept of loop computing for solving matrix equations and accelerating ML algorithms in one step, hence this technology speeds up the training phase for cloud and edge computing, thus saving time and cost for the data centres owners, and enabling low-power artificial intelligence (AI) processing at the edge.

Campo di applicazione

<p>• Big data analysis;</p> <p>• Machine learning.</p>

Vantaggi

<p>• In-memory analogue computation, where data analysis is performed directly in analogue memory with no need to transfer data;</p> <p>• Physical matrix-vector moltiplication within a XP circuit thanks to Kirchoff’s Law and Ohm’s Law;</p> <p>• Physical iterqation thanks to the feedback loop instead of time-consuming numerical iteration.</p>

Stadio di sviluppo

Simulations and laboratory prototipe

Contatto

licensing.tto@polimi.it