Neuromorphic Chip Biz Heats Up

Old concept gets new attention as device scaling becomes more difficult.


It’s no secret that today’s computers are struggling to keep up with the enormous demands of data processing and bandwidth, and the whole electronics industry is searching for new ways to enable that.

The traditional approach is to continue to push the limits of today’s systems and chips. Another way is to go down the non-traditional route, including an old idea that is generating steam today—neuromorphic computing.

Originally conceived by engineering guru Carver Mead in the 1980s, neuromorphic computing and its previous incarnation, artificial neural networks, make use of specialized chips that are inspired by the computational functions of the brain. Neuromorphic technology, sometimes called brain-inspired computing, is a paradigm shift that breaks away from Moore’s Law. Neuromorphic chips don’t require costly leading-edge processes.

In simple terms, neuromorphic chips are fast pattern-matching engines that process the data in the memory. In theory, these chips promise to enable systems that can perform several tasks, such as computer vision, data analytics and machine learning. The ultimate goal is to realize true artificial intelligence (AI).

Today, Facebook, Google and others are handling many of these intelligent-like tasks using traditional computers and chips. In this topology, sometimes called the von Neumann architecture, the system has three main components—a processor, memory, and storage. They are connected via a systems bus.

The industry, however, is running into an I/O bottleneck with today’s systems, at least for many intelligent-like applications. So for these apps, the industry hopes to develop a new class of neuromorphic systems and chips, although they won’t replace traditional technology for the foreseeable future.

“For many problems going forward, (von Neumann hardware) will still be the right solution,” said Geoffrey Burr, a principal research staff member at IBM Research. “But there’s an enormous amount of work that needs to be done to make those (intelligent-like) algorithms work in software on regular von Neumann hardware. The problem is that you need this steady stream of data through the bus. So, you’re spending a lot of energy and time shipping that data in and out.

“It would be ideal to bring the computation to where the data is,” Burr said. “That’s where we see the opportunities for these neuromorphic systems. It will accelerate machine learning.”

For that reason, neuromorphic technology is finally heating up after years of research under the radar. Until recently, General Vision was one of the few vendors shipping neuromorphic chips. But in a move that could propel the technology, General Vision recently licensed its intellectual-property to Intel, which is shipping embedded processors based on the technology.

In addition, IBM recently launched TrueNorth, a 1 million-neuron processor. Meanwhile, a European consortium, as well as HP, Qualcomm, Samsung and several chip startups, are also pursuing the technology. And several universities and government agencies, such as the U.S. Department of Defense (DoD), are working on it.

Still, neuromorphic technology faces an uphill battle to gain broad adoption. Some vendors are shipping neuromorphic chips, but others are struggling. It’s difficult to replicate the functions of the human brain in silicon, and the industry’s understanding of the brain works remains limited.

In addition, neuromorphic chips require a different way of programming the data. Plus, the current chips may need a memory overhaul. All told, the neuromorphic chip market is a small business today, but it is expected to grow at an annual rate of 26% and reach $4.8 billion by 2022, according to Markets and Markets, a publisher of research reports.

AI winter
The field of neuromorphic technology and its previous incarnation, artificial neural networks, was a hot market in the 1980s. At the time, several companies were pursuing this and other technologies to enable AI and so-called expert systems. An expert system is a computer that mimics the decision-making ability of a human.

“Back in the 1980s, everybody thought expert systems were going to take over the world,” said Dave Schubmehl, an analyst at International Data Corp., a market research firm. “But we ended up having what we call an AI winter. We really didn’t have enough data. We really didn’t have enough compute power to make these expert systems actually useful.”

Needless to say, compute power has dramatically increased over the years, but the amount of data also has exploded.

Nevertheless, thanks to the increase in compute power, Facebook, Google and others have developed an AI-like technology called machine learning or deep learning. This technology makes use of software algorithms, which in turn can learn and make predictions from various data.

A subset of this technology is called unsupervised machine learning. This makes use of artificial neural networks to crunch the data. “Essentially what happens is that the neural network algorithms crunch on the data long enough to identify patterns and identify a set of attributes associated with those patterns. Over time, it learns which of those attributes are important,” Schubmehl said.

“A lot of these applications are able to run effectively on (traditional) GPUs,” he said. “The question is if we take that to a non-von Neumann architecture, can we potentially get a quantum leap? It’s really too early to tell if (neuromorphic technology) can really be applied on a broad scale.”

Neuromorphic technology is appealing, though. “The von Neumann architecture is more like serial execution of the instructions. You access the memory from outside and somewhat closer to the CPU,” said Srinivasa Banna, a fellow and director of advanced device architecture at GlobalFoundries. “With neural networks, you can do things in parallel with efficient elements. Those are all energy efficient. But it requires innovation in the way the data is processed in parallel.”

What are neuromorphic chips?
To some degree, the industry has been shipping neuromorphic chips for select markets, particularly for pattern matching applications. Here are some hypothetical examples of the possible apps:

Military systems. Using neuromorphic technology, a drone could identify and match potential targets on the battlefield. It also could learn new data in-flight.
Facial recognition. With a camera and neuromorphic chips, a system could accurately match images in real time.
Industrial gear. Using the technology, a camera-enabled system could find small defects in chips in the fab without the need of today’s wafer inspection equipment.

“Today, computing is not good for technologies like parallel pattern recognition,” said Guy Paillet, chief executive of General Vision, a supplier of neuromorphic chips. “I am not saying a computer, a CPU, a GPU and all can’t do it. They are not very efficient.”

Indeed, neuromorphic technology is different than traditional chips. According to Stanford, there are two classes of neuromorphic computing—artificial neural networks, and biology-based learning models that mimic the brain.

To some degree, chips based on the neural network approach have gained traction. In contrast, chips based on the biology model are still trying to get off the ground, as the industry has hit some roadblocks in the arena.

In both approaches, neuromorphic chips consist of multiple neurons and synapses. These aren’t biological neurons and synapses, but they mimic the functions of these structures.

For example, General Vision’s neuromorphic chip consists of 1,024 neurons, all interconnected and working in parallel. A neuron is a 256-byte memory based on SRAM, plus 3,000 logic gates.

Multiple chips can be daisy-chained in a network, forming the basis of an artificial neural network. A network could have a multitude of neurons and synapses. The synapse connects one neuron to another.

Basically, an artificial neural network has three layers—input, hidden, and output. In operation, a pattern is first written in a neuron in the input layer. The pattern is broadcast to the other neurons in the hidden layers. Each neuron reacts to the data. Using a weighted system of connections, one neuron in the network reacts the strongest when it senses a matching pattern. The answer is revealed in the output layer.

A neural network also consists of a learning mechanism. The weights of the connections are modified based on the input patterns.

And unlike today’s chips, neuromorphic devices conduct the processing in the memory. This enables faster processing, but the current chips are based on SRAM. SRAMs are fast, but they are power-hungry and take up too much area. So neuromorphic chipmakers hope to move from SRAM toward a next-generation memory type, such as phase-change or ReRAM. In addition, phase-change and ReRAM promise to enable spike-based learning techniques in chips. In biology, neurons send messages to each other via precisely-timed pulses or spikes.

The neuromorphic community also wants to bring spiking, or a timing element, into their chips, but this is one of the major roadblocks in the industry. “From an engineering and electrical perspective, spike coding can be more complex to implement,” said Christian Gamrat, a researcher from CEA-Leti.

For its part, CEA-Leti is developing a memristive-based device array for use in spike-based coding. The array is based on a 1T-1R CBRAM technology. “Although this is a work in progress, we believe that the combination of memristive technologies and spike-based coding is a promising way for the efficient implementation of embedded neuromorphic architectures,” Gamrat said.

In any case, after years of promises, the industry is finally showing results on several fronts. On one front, General Vision for some time has been shipping 130nm neuromorphic chips, mainly for industrial and related applications.

General Vision also licensed its IP to Intel, which is using the technology in its so-called Curie Module. The module features a 32-bit Quark SE embedded processor as well as General Vision’s pattern-matching IP.

The module could propel neuromorphic technology into some broader markets. “Curie’s dedicated sensor hub can be used in a variety of ways,” said Brian Krzanich, chief executive of Intel, at a recent event.

The module is targeted for fitness trackers and other wearables. General Vision’s IP optimizes the analysis of the sensor data in systems, enabling fast identification of actions and motions.

Others are making progress in more traditional computing applications. In 2013, for example, the European Union launched the Human Brain Project, an effort to gain a better understanding of the brain. As part of this effort, the project is developing two neuromorphic computing platforms.

The goal of these projects is to accelerate machine learning times in high-end systems. One project, dubbed SpiNNaker, is a massively-parallel computer architecture based on 1,036,800 ARM9-based cores.

Meanwhile, the second project, called BrainScaleS, aims to develop a 180nm chip with 512 neurons and 131,072 synapses. The BrainScaleS project also involves a chip based on mixed-signal technology. “A mixed-signal approach combines space and energy efficient local analog computing with the scalability of a binary system,” said Karlheinz Meier, a co-leader of the project.

Meanwhile, IBM is going down the digital path. Last year, IBM rolled out TrueNorth, a chip that consists of 1 million neurons, 256 million synapses and 4,096 parallel cores. The 5.4 billion transistor device is based on SRAM and a 28nm process from Samsung.

In addition, IBM is also exploring the idea of migrating from SRAM to phase-change memory. This architecture is potentially 25 times faster at lower power than systems based on traditional GPUs, according to IBM. “It’s going to be difficult for any nonvolatile memory element to have the performance of SRAM. There will always be a place for SRAM, but SRAM is area hungry,” IBM Research’s Burr said. “One of the nice things with phase-change is that it has a huge range of resistive states, from very resistive to quite conductive. So when you want it as an analog element, it’s very attractive.

“By performing computation at the location of data, non-von Neumann computing ought to provide power and speed benefits. For one such non–von Neumann approach, on-chip training of large-scale artificial neural networks using nonvolatile memory-based synapses viability will require at least two things,” Burr said. “First, despite the inherent imperfections of nonvolatile memory devices, such as phase-change memory or resistive RAM, (they) must achieve competitive performance levels versus artificial neural networks trained using CPUs or GPUs. Second, the benefits of performing computation at the data must confer a decided advantage in either training, power or speed or preferably both.”

In any case, the question is clear: “Where is the technology heading?”

“Machine learning exists today,” Burr said. “We realize that it’s not enough. We would like to have machines that learn from their own experiences. Beyond that, we are at the beginning of transcending from machine learning and driving towards machine intelligence.”

For this, the industry could go in several directions. In the near-term, companies will continue to use today’s systems.

Long term? The industry may use traditional chips. “There is also a possibility that technologies like TrueNorth will lead to things that are non-von Neumann hardware,” Burr said. “That may accelerate machine learning. Five years from now, we will know whether it will happen or not.”