Released on: January 16, 2008, 11:10 pm

Press Release Author: Recognetics.Co.,Ltd.

Industry: Electronics

Press Release Summary: CM-1K can be used for machine vision, video analytics, speech
recognition, signal recognition or data mining

Press Release Body: New neural network chip adds intelligence to smart sensors

Jan.17, 2008 Suzhou, China -

Recognetics is pleased to announce the availability of CogniMem1K (CM-1K), a neural
network chip featuring 1024 neurons working in parallel, and which can be daisy
chained to other CM-1K chips to increase the network size. It is an ideal companion
chip for smart sensors and cameras and can classify patterns at high speed while
coping with ill-defined data and unknown events, and adapting to changes of contexts
and working conditions. Depending on the source of the input patterns, CM-1K can be
used for machine vision, video analytics, speech recognition, signal recognition or
data mining. "CM-1K is a component which has been long awaited by researchers and
industries interested in using neural networks to solve real problems. Its
performance and affordability should make it an essential enabler for artificial
intelligence in our everyday life" says Wo Lin, president of Recognetics.

CM-1K implements two powerful non-linear classifiers (RCE and KNN) in a natively
parallel architecture. The tremendous benefit of this architecture is a recognition
cycle that remains under 11 microseconds regardless of whether the entire network is
composed of one, two or many chips. Brute computational power is equivalent to 27.3
giga operations/second @ 27 MHz for a single chip, twice as much or 54.6 giga
operations/second for two chips, and so on. \"The CM-1K represents a major
breakthrough in pattern recognition that leverages neural networking technology\",
said Jamshed Qamar, Vice President of Customer engineering at ChipX, the company
hired for the ASIC design. \"We faced many challenges in implementing this design
including the integration of more than 10 million gates, managing power distribution
and maintaining signal integrity throughout the chip with 1024 neurons all switching
at the same time. Maintaining close timing, with all neurons communicating with
each other in less than half a cycle, and enabling multiple chips to be cascaded
without impacting performance, further added to the challenge\". High-speed and
trainability make CM-1K a practical solution for real-time, distributed, intelligent
devices in industrial automation, robotics, security, health monitoring, predictive
maintenance, intelligent transportation and more. Also stackability is critical for
data mining applications in bioinformatics, medical imaging, satellite imaging, data
center management and more.

Recognetics offers a line of evaluation boards for developers and OEMs.
The CM-EB, is a base board featuring an Actel FPGA, a USB interface and one CM-1K
chip. It allows to evaluate the two classifiers available in CM-1K, the trainability
of the neurons, and the knowledge bases built by the neurons. It can also be used to
test the optional engine for the recognition of vectors received on the CM-1K
digital input bus. Developers can add pre and post-processing functions in the FPGA
to condition or generate the vector data and to consolidate or report the results.
CM-EB is supplied with a development library and control panel. Additional CM-1K can
be added to the board.

A second board, CM-IR, is intended for image recognition delivering up to 60
recognitions per second. It features a Micron CMOS sensor, an Actel FPGA and one
CM-1K chip. The video signal is entered directly into the CM-1K digital input bus
and a module internal to the chip sub-samples the pixel values into a signature
vector. Alternatively, programmers can choose to implement their own signature
extraction in the FPGA. In either case, the vector is broadcasted to all of the
neurons and the response of the neuron with the best match is available in 11
microseconds. This data can be transmitted over GPIO lines, RS232 or I2C bus. As an
option, programmers can condition the results of the recognition in the FPGA prior
to transmission. The neurons can be trained in real-time, or a knowledge base can be
loaded from a file and saved to a Flash memory so the board resumes recognition
autonomously at the next power up. CM-IR can be used for industrial inspection,
machine vision, video surveillance, robotics and more. It is delivered with a
development library and EasyTrainer software. Additional CM-1K can be added to the
board.

Recognetics is the leading provider for fully parallel, high performance pattern
recognition semiconductors based on the CogniMem neural network patented technology.
Recognetics supplies standard and custom modules designed around CogniMem and
suitable for signal processing, speech and image recognition, data mining. For more
information about Recognetics, visit http://www.recognetics.com.


Web Site: www.recognetics.com

Contact Details: 80# Xiangyang road,Suzhou new district, China
Tel:86-512-68411183
Fax:86-512-68085081
E-mail:dsn@recognetics.com.cn
Contact person: Ms.Dai