Search

link to homepage

Advanced Computing Architectures (ACA)
towards multi-scale natural-density Neuromorphic Computing

Navigation and service

 

The overall goal of the project ACA is to develop neuromorphic computing modules which are specifically designed to address neuroscientific questions and which shall be integrated in traditional supercomputing system. Partners from five different institutes at Forschungszentrum Juelich, from RWTH Aachen University, from the University of Heidelberg and the University of Manchester work together in this project which is funded by the Helmholtz Initiative and Networking Fund.Neuromorphic computing refers to advanced computing architectures exploiting principles underlying the superior performance of the brain: in everyday tasks such as object recognition in natural environments, brains outperform traditional (von-Neumann) computers in terms of computational capacity, robustness, processing speed and energy efficiency. Current challenges in this field are the realization of the complex high-density connectivity of the brain, the resulting communication between network elements, the plasticity of connections and the problem of fast network instantiation in the computing device. ACA aims at overcoming these challenges by assuming a network-centric view: it does not start with the design of the smallest computational elements, but with the analysis of the requirements imposed by the instantiation and ongoing modification of brain-scale networks and the communication therein. The project is thereby targeting a breakthrough: once neuromorphic systems can cope with natural-density networks, connectivity is no longer a barrier for any brain-like computation.

Test cases are selected from neuroscience because of the fundamental limitations traditional approaches face in this area. Developing a synthetic neuromorphic system by addressing neuroscience questions ensures that the design remains generic and compatible with the principles of nature.

The consortium is distinguished by a broad range of competences from neuroscience, over semiconductor physics to circuit design and system integration to model description languages up to applications. Furthermore, it includes experts with a decade of experience in prominent European neuromorphic-computing initiatives, and excellent access to a user community. The partners with existing neuromorphic systems and the developers of simulation code for traditional computers ensure that these systems and the community rapidly and sustainably profit from the results.

Main challenges and concepts

In the following, we summarize the main challenges of this project and briefly describe the concepts we will use to face these challenges:

1) Account for the fact that the critical biological ingredients underlying high brain performance are unknown

  • tight interaction between Neuromorphic Computing and Neuroscience
  • continuously track and incorporate progress in Neuroscience
    synergies:

    • identification of biological computing principles (concepts, constraints)
    • NC platform for neuroscience simulation (e.g. studies on plasticity and learning)
    • applications (science cases) and benchmarks (test cases)
  • flexibility-efficiency trade-off

    • adaptability to new models
    • accuracy and reproducibility (focus on deterministic approaches)
    • exploration of non-deterministic approaches for future NC architecture (see Task 8, Task 10, Task 21)

2) Realization of high-density (~105 neurons/mm3 , ~104 synapses/neuron), multi-scale (short- and
    long-range interactions), heterogeneous (non-stereotype), plastic (short- and long-term + structural
    plasticity) connectivity, and communication therein

  • network-centric view
  • regard network instantiation, adaptation and communication as central problem to be accounted for from the beginning
  • integration of new technologies and devices (resistive switches)
  • scalability

3) Accelerated dynamics to enable research on plasticity, learning and development

  • flexibility-efficiency trade-off
  • efficient accelerators for frequent and computationally complex operations
  • modularity (neuromorphic accelerators) and integration in traditional HPC infrastructure  

4) System integration, maintenance & community integration

  • modularity: integration in traditional HPC infrastructure
  • rather than standalone hardware system
  • early arrive at products that can be delivered to the community (short iterations)
  • co-development of neuromorphic systems and HPC simulation technology
  • testing of neuromorphic solutions against state-of-the-art HPC
  • traditional HPC benefits from new neuromorphic concepts
  • address established user community
  • existent user community for BrainScaleS, SpiNNaker & NEST
  • shared software interfaces (e.g. PyNN) and standardized languages

Servicemeu

Homepage