Related Experiment Videos
Multiprocessor and memory architecture of the neurocomputer SYNAPSE-1
1Siemens AG, Corp. R & D, Munich, Germany.
International Journal of Neural Systems
|December 1, 1993
Summary
The SYNAPSE-1 neurocomputer offers flexible, high-speed neural network processing. Its novel architecture significantly accelerates computations, outperforming standard workstations by 8000 times.
Area of Science:
- Neurocomputing
- Computer Architecture
- Artificial Intelligence
Background:
- Existing neurocomputers often lack flexibility in adapting to diverse neural algorithms.
- Scalability in processing power and memory is crucial for accommodating varied application demands.
Purpose of the Study:
- To introduce SYNAPSE-1, a general-purpose neurocomputer with a flexible multiprocessor and memory architecture.
- To demonstrate the significant speed-up capabilities of SYNAPSE-1 for neural algorithms, including learning.
Main Methods:
- Implementation of a 2-dimensional systolic array of neural signal processors (NSPs).
- External storage of weights, allowing independent scaling of memory and processing power.
- Development of a dedicated neural algorithms programming language embedded in C++.
Main Results:
- SYNAPSE-1 exhibits a speed-up factor of several orders of magnitude for neural computations.
- A prototype demonstrated an 8000x speed improvement over a standard workstation in benchmark tests.
- The architecture allows for individual adaptation of memory size and processing power.
Conclusions:
- SYNAPSE-1 provides a highly flexible and powerful platform for a wide range of neural algorithms.
- The neurocomputer architecture achieves substantial computational acceleration, particularly in learning tasks.
- The design facilitates customized configurations for specific application requirements.