Startup’s pattern matching tech is inside Intel’s Quark neural net
The Quark SE is the system-chip on Intel’s button-sized Curie module that was launched at the Consumer Electronics Show. Intel said at the time that the QuarkSE chip, a processor developed for wearable applications, included a pattern classification engine that allows it to identify different motions and activities.
"Yes, the pattern matching/classification engine inside the Quark SE is an implementation of our technology," Philippe Lambinet, CEO of NeuroMem told eeNews Europe in email. But Lambinet said he could not say where Intel had licensed the technology from, but that it was not from NeuroMem as the company had only been incorporated recently.
The pattern matching engine inside QuarkSE and offered by NeuroMem has had a convoluted history.
Philippe Lambinet, CEO of NeuroMem
The technology as originally developed by General Vision Inc. (Petaluma, Calif.) a company founded in 1987 and led by expatriate French experts in computer vision and neuromorphic computing. Guy Paillet, CEO of General Vision, developed the ZISC (Zero Instruction Set Computer), a silicon neural network chip with IBM (France) in 1993. He joined General Vision in 1999. Anne Menendez, CTO of General Vision, and Paillet started developing the CM1K, a neuromorphic pattern-recognizing chip with 1,024 neurons between 2006 and 2008. This was originally manufactured as a 130nm CMOS chip by Oki Semiconductor.
In 2011 General Vision created a subsidiary company, CogniMem Technologies Inc. (Folsom, Calif.), to market and sell the CM1K chip. However, General Vision’s involvement with CogniMem was terminated in 2013. CogniMem carried on as a reseller of the CM1K chip while General Vision continued to hold the rights to the intellectual property. In 2014 General Vision renamed its technology NeuroMem and created NeuroMem Inc. with Lambinet, formerly a corporate strategy officer at STMicroelectronics NV, as chief executive.
Next: The Intel example
Lambinet said that although the Intel deal was not made with NeuroMem, things have been changed around so that now NeuroMem is able to do licensing of the technology. "General Vision now focuses on turn-key vision systems and is one of NeuroMem’s customers," Lambinet said in email.
"We can license our IP to semiconductor companies for integration in their SoC or we can license to OEM customers doing their own SoCs or FPGAs. NeuroMem also sells standard ICs, boards and development tools that our customers use to build their systems," Lambinet said.
Lambinet said the Curie module is a great example of NeuroMem’s technology as it takes signals from an accelerometer and gyroscope on the board and "understands human physical behavior that it has learned to recognise."
He added that up until now sensor-based peripherals had to be connected to a smartphone or to the cloud to perform any useful classification. With the NeuroMem technology deployed at the end-point node the sensing unit can become autonomous and does not need to consume bandwidth and power to transmit unfiltered data.
The NeuroMem technology can also be deployed elsewhere in the network because it scales well, Lambinet said. "We recognise one face out of millions just as fast as we recognise one face out of 1000. CPU/GPU solutions are extremely fast for small and medium datasets but they slow down dramatically as soon as they reach the limit their parallelism. At some point, they become sequential as all the cores have to share memory bandwidth. Our technology does not have this limitation because the computing happens inside the memory."
Next: More about the CM1K
The CM1K is a chain of 1,024 identical neurons that operate in parallel but are connected together to make global decisions. A neuron is 256bytes of SRAM with some associated programmable logic to compare an incoming pattern with a reference pattern held in memory and to "fire" according to its similarity.
Collectively the neurons behave as a K nearest-neighbor classifier or an implementation of a radial basis function and are trainable. They are well-suited to working with ill-defined and fuzzy data and a high variability of context.
A single CM1K matching a pattern of 256 bytes against 1,024 examplars can deliver the equivalent of 192 giga operations per second, according to the CM1K datasheet. The chip operates at a frequency of 16 MHz to 27 MHz and dissipates about 0.5 watts.
NeuroMem: www.neuromem.com
General Vision: www.general-vision.com
Related news articles:
Rousset judgment is economic and social mistake
Plan to re-open Rousset fab requires investors
NeuroMem IC Matches Patterns
If you enjoyed this article, you will like the following ones: don't miss them by subscribing to :