MENU

NeuroMem IC matches patterns, sees all, knows all

Technology News |
By eeNews Europe

If you’ve ever seen the U.S. TV series “Person of Interest,” during which an anonymous face in the Manhattan crowd, highlighted inside a digital frame, is identified by “The Machine” as the victim (or perpetrator) of an impending crime, you get the picture: pattern recognition, the heart of ‘Big Data’ analytics, is part of everybody’s future.

Scene from 'Person of Interest'

Scene from ‘Person of Interest’

The TV series, although a fictional drama, foreshadows a real-life escalation in the era of massive data collection.

IoT, Big Data or Big Brother, “the buzzwords are all about data collection and data interpretation,” Philippe Lambinet, former STMicroelectronics executive vice president, told us in a recent interview in Europe.  

Analytics is necessary so that [a machine] can make relevant decisions. “Data collection is becoming so enormous that we need new ways of doing data analytics,” he said.

Lambinet, now the acting CEO of a startup called NeuroMem, is betting his farm on the growing demand for pattern recognition in data analytics.

Unlike established computing architectures, which Lambinet thinks are too slow and too power hungry for pattern matching, NeuroMem, a fabless chip company, is promising to enable efficient pattern recognition by bringing massively parallel, low-power, real-time, cost-effective computing architecture with a built-in capability to learn on the fly.

Comparing two pattern-matching architectures: Sequential vs. Parallel
(source: NeuroMem)

Comparing two pattern-matching architectures: Sequential vs. Parallel

(source: NeuroMem)

 

Most people implementing pattern matching algorithms use the traditional computing model also called the von Neumann architecture, said Lambinet.

They implement, in software, a fundamentally parallel algorithm and then they run it on a fundamentally sequential hardware. He thinks the need for analyzing massive data is changing that landscape and creating serious demands for non-von Neumann computing architecture.

Lambinet believes the startup will be able to raise enough money — about $5 million — by the end of this month to get its business going. 

Five mil doesn’t sound like a lot of money for a fabless semiconductor startup.


But the nice thing about NeuroMem, said Lambinet, is that “we are leveraging the work already done by General Vision.” General Vision, a Petaluma, Calif. company founded 27 years ago, has long specialized in image processing technology for machine vision, and image recognition systems based on neural networks.

In short, NeuroMem already has a proof of concept for its massively parallel computing architecture needed for pattern recognition, its intellectual property and even a commercial chip to boot.

General Vision’s CEO Guy Paillet and its founder and CTO Anne Menendez are major shareholders and board members of the startup. NeuroMem’s charter is to develop subsystems using a General Vision-developed chip called CM1K (which was built on a 130nm process technology), design the successors to CM1K at a finer geometry node, and promote them for vision and non-vision applications which include text and data analytics, audio and sound recognition and secure networking. General Vision, which will carry on its business as usual, has no plans to enter the chip business, said Lambinet.

Pattern matching chip
NeuroMem’s pattern-matching chip consists of a chain of identical neurons. Each neuron has the same behavior. These neurons are interconnected via a small bidirectional bus. All are designed to receive and execute the same pattern-matching/classification calculation in parallel. The technology is scalable, offering the same behavior independent of the size of network.

Philippe Lambinet

Philippe Lambinet

More specifically, each neuron consists of SRAM and a small programmable logic unit. The logic unit, not a fully programmable CPU, is prewired to run certain types of algorithms. When those neurons are chained together, they run entirely in parallel. “Eighteen clock cycles later, the result is the closest match to the original pattern,” explained Lambinet. “You can run a million of neurons or a billion of them together.”  The very architecture of a neuron removes the memory bottleneck, since it blends pattern memory with pattern learning and recognition logic.

The pattern-matching chip works in real time.  “In the order of a few micro-seconds, it provides constant recognition, classification and learning latencies,” the company said.

It’s also energy-efficient. 1,000 neurons (implemented in existing silicon CM1K, for example) that deliver 100+ Giga Operations per second “consumes 0.5 watts at 27MHz,” according to NeuroMem.

Intel’s Curie
Nothing comforts Lambinet more than knowing that NeuroMem isn’t the only company engaged in pattern-matching acceleration technologies. 

He pointed out, “Look no further than Intel’s announcement last month at the International Consumer Electronics Show.” Intel’s new platform, called “Curie,” is designed for companies interested in developing wearable technology.

At its core is Intel’s first wearable SoC that can run for extended periods from a coin-sized battery. It features a motion sensor, Bluetooth radio, and battery charging capabilities. Inside Curie is a low-power integrated DSP sensor hub “with a proprietary pattern matching accelerator,” according to Intel. Intel appears to see the need for “a pattern-matching accelerator,” to analyze the data provided by its sensors and classify different human activities (running, jumping, walking, etc.).

When asked by EE Times whose “proprietary pattern-matching accelerator” Curie is using, the Intel spokeswoman told us, “We are not disclosing additional details at this time.”

Regardless whose pattern- matching technology Intel is using, one thing is clear in Lambinet’s mind. Intel, the CPU giant, isn’t using a CPU for pattern matching. CPU aren’t low power enough, and not fast enough, “for this sort of task,” he noted.  

Neural network on the rise
It’s important to note that in recent years, advanced scientific fields such as neural network and deep learning are spreading across key players including Facebook, Google, Microsoft and Baidu. These firms have been hiring specialists, presumably to build web services that automatically understand natural language and recognize images.


A few chip vendors are also dabbling with neuromorphic computing, some seeking to solve data analytics problems while others try to reproduce, on silicon, the way the human brain works, explained Lambinet. IBM is working with SyNAPSE chips, NVidia is pushing brain-like computing with new graphics products based on thousands of CUDA cores,and Qualcomm is reportedly working on neuro-inspired chips

Genesis: IBM ZISC chip
But the brain chip – more advanced in terms of commercial availability – is already here, said Lambinet, in that old CM1K chip developed by General Vision.

The genesis of the CM1K chip, however, is actually in Paillet. He was a co-inventor of the IBM ZISC (zero instruction set computer) chip. Lambinet called the IBM chip “the very first hardware pattern-matching accelerator ever produced.”

The IBM ZISC chip was a neural network chip invented and patented by a team of engineers at IBM France and Guy Paillet in 1993. IBM manufactured it from 1993 to 1999. As IBM was ending chip production in France, the ZISC chip’s successor, CM1K, emerged at General Vision. It was released in August 2007.

General Vision has already proven its pattern recognition technology by using its CM1K and powering vision systems with real-time learning and recognition. By stacking NeuroStack boards (starting with 4,096 neurons up to 1,000,000 for a stack of 250 boards), it’s possible to “check 100,000 fingerprints per second against a database of one million prints,” or “check 100,000 iris scans per second against a database of one million eyes,” according to General Vision.

General Vision’s neuromorphic technology based on CM1K has been licensed and used prominent institutions around the globe, including laboratories of the U.S. Air Force and the French Army, according to Lambinet. The Indian government, meanwhile, is also known to have licensed General Vision’s pattern matching technology, but not the CM1K chip itself. They are running the General Vision’s algorithms on FPGA to identify people. “We know this works very well,” said Lambinet.

NeuroMem’s strategy
NeuroMem will be updating the CM1K chip as soon as it gets necessary funding. The current CM1K chip comes with 1024 neurons per chip, with each neuron memory at 256 bytes. Because General Vision isn’t a chip company and never wanted to be, the front-end ASIC design for its CM1K was done by Oki Semiconductor, which later became Lapis Semiconductor. The CM1K was fabricated by UMC in Taiwan using a 130nm process. 

For the time being, though, NeuroMem is rolling out what it calls a “BrainCard.” It combines the CM1K pattern-learning and pattern-matching capability (up to 9,000 neurons) with vision and audio sensors. It has connectors for Raspberry Pi, Arduino and PMOD, rendering the extension capability virtually unlimited, Lambinet explained.

Braincard, frontside, with CM1K expansion stacked
(source: NeuroMem)

Braincard, frontside, with CM1K expansion stacked

(source: NeuroMem)

It can also integrate an Intel Edison, making the BrainCard “a standalone pattern recognition machine, with a very simple programming environment, the size of a credit card,” he said. “And its pattern recognition processing speed can be compared to a super computer, without the kW of power consumption, of course.”

Braincard, backside, with Intel Edison mounted
(source: NeuroMem)

Braincard, backside, with Intel Edison mounted

(source: NeuroMem)

Although NeuroMem initially launched the BrainCard on Indiegogo, Lambinet acknowledged that its target market isn’t exactly the so-called maker crowd. “We did it to create buzz among big corporations and VCs.”

The capability of the BrainCard is eerily similar to the fictional “Machine.” Lambinet listed, as its potential applications, face recognition (one face plucked from several thousands in a few microseconds), ECG/EKG signal recognition to detect anomalies, industrial visual inspection, voice recognition, and others. “Basically any pattern that can be digitized, can be learned and then recognized,” he said.

But NeuroMem, being a startup, needs a focus to market its products and technology. Asked his strategy, Lambinet said industrial vision (not crime-stopping) is the priority.  “We regard it as a sweet spot.”

He pointed out that of the 1.5 million workers employed by Foxconn, for example, many are actually doing visual inspection. It’s a huge market that can be automated. “Visual inspection is very important because it touches quality.”

Second, Lambinet foresees the use of NeuroMem’s pattern matching in big data analytics. Today, racks of GPU boards are used in data centers, although they aren’t designed for parallel computing. NeuroMem is confident that it can go after the same market.

Third, NeuroMem plans to go after “big guys” for licensing, Lambinet said. Noting that many companies such as mobile handset giants like Samsung and sensor companies like Bosch, STMicroelectronics and InvenSense are working on SoCs and vision/motion sensors, “We want those guys to produce billions of chips using NeuroMem’s technology,” he added.

— Junko Yoshida, Chief International Correspondent, EE Times


Share:

Linked Articles
eeNews Europe
10s