Everspin MRAM Powers Next-Gen CogniBlox Reconfigurable Module

Offering a combination of high performance, endurance, non-volatility and reliability, magnetic RAM from Everspin Technologies, Inc., is the memory solution for a new cognitive computing platform.

CogniMem Technologies, Inc. is using MRAM to quickly store and restore the content of the neurons on its recently introduced CogniBlox reconfigurable module, which features four CM1K neuron chips or a total of 4,096 silicon neurons that mimic how the human brain processes data.

MRAM offers SRAM-like speeds, non-volatility and unlimited endurance, making the technology suited for massively parallel pattern recognition tasks used in cognitive computing. MRAM’s nanosecond-class access times provide the performance necessary to support real-time learning and pattern recognition. The CogniMem silicon neurons feature an automatic model generator, which enables them to learn examples and build and store knowledge in real time. MRAM is non-volatile, meaning it saves the knowledge in the case of a power interruption and without the use of external components such as batteries.

Another usage of the MRAM on the CogniBlox platform is the ability to save and restore diferent knowledge files into the neurons while an application is running. For example, neurons can be loaded with a primary knowledge recognizing the silhouette of a person in a landscape and if they recognize a human presence in the field of view, the neuron content can be modified to hold a secondary knowledge and identify who it is, where he is looking, etc.

CogniBlox is a stackable and reconfigurable module allowing the design of versatile, massively parallel pattern recognition architectures for high-performance cognitive computing, sensor fusion, video analytics and more. Applications span machine vision and speech recognition to data mining. Each CogniBlox board uses 4MB of MRAM to store the knowledge files (‘learned models’), as well as buffer input data such as signature vectors, video images and more. The board is reconfigurable through a Field Programmable Gate Array (FPGA).

"CogniMem has a unique chip technology that implements high speed and parallel pattern recognition in hardware, and MRAM is ideally suited to deliver the performance and endurance desired for this cognitive computing application," said Bruce McCormick, co-founder, president and CEO of CogniMem. "Our goal is to demonstrate a network of one million neurons, a capacity where high speed save and restore is mandatory and that will require MRAM if the power fails or is turned off."

CogniBlox enables system configurations of up to one million silicon neurons using 250 boards that deliver unprecedented performance of 0.13 petaops with a typical power consumption of 500 watts, allowing for 256 million connections every 10μs.

"CogniMem is at the leading edge of cognitive computing and we are excited that MRAM is the only memory that can enable real-time cognitive computing," said Phill LoPresti, Everspin president and CEO. "Everspin MRAM continues to enable revolutionary products through its unique combination of working memory and storage characteristics offering a new storage class memory."

Everspin will be showcasing the CogniMem CogniBlox module at Embedded World in the Nuernberg Convention Center, Germany, Feb. 28 – March 1.