CUBRIC - Skräddarsyd hjärninspirerad integrerad krets
|Koordinator||KUNGLIGA TEKNISKA HÖGSKOLAN - Skolan för informations- och kommunikationsteknik, Avdelningen för elektroniksystem|
|Bidrag från Vinnova||1 513 180 kronor|
|Projektets löptid||januari 2014 - december 2016|
Syfte och mål
The overall objective of CUBRIC is to develop and demonstrate a custom hardware implementation of a scalable brain-inspired computational architecture based on neural network. The project not only fulfils the brain-inspired hardware development aim but also has covered different hardware optimizations during the hardware development. As one of the key goals was to help the Applicant developing his career, right after the project, he got an offer for senior lecturer from Mälardalen University to continue working on the same area.
Resultat och förväntade effekter
The obtained results show that under the same technology process, the mixed signal implementation of brain-inspired architecture consumes considerably less power and silicon area than that of the state-of-the-art processor core. Our evaluation show that CUBRIC can manage the multicast-based traffic of neural networks better than the other topologies used in many neural network implementations which directly translates to faster neural processing. This is achieved through the proposed reconfigurable and customizable communication fabrics in neural-based systems.
Upplägg och genomförande
The overall technical contribution of CUBRIC are briefly summarized here: 1. Hardware realization of both digital, analog, and mixed signal (analog and digital) implementations of spiking neural network. 2. Reconfigurable and customizable communication fabrics to support the heavy inter-neuron communication traffic of neural networks in a very efficient way. 3. Developed a framework to interleave spiking neural networks on a coarse grained recongurable architecture (CGRA), which made significantly power and performance gains.