Specialized AI Chips Hold Both Promise and Peril for Developers

Specialized AI Chips Hold Both Promise and Peril for Developers

Trending Technologies   August 12, 2019

When it comes to the compute-intensive field of AI, hardware vendors are reviving the performance gains we enjoyed at the height of Moore’s Law. The gains come from a new generation of specialized chips for AI applications like deep learning. But the fragmented microchip marketplace that’s emerging will lead to some hard choices for developers. 

The new era of chip specialization for AI began when graphics processing units (GPUs), which were originally developed for gaming, were deployed for applications like deep learning. The same architecture that made GPUs render realistic images also enabled them to crunch data much more efficiently than central processing units (CPUs). A big step forward happened in 2007 when Nvidia released CUDA, a toolkit for making GPUs programmable in a general-purpose way.

AI researchers need every advantage they can get when dealing with the unprecedented computational requirements of deep learning. GPU processing power has advanced rapidly, and chips originally designed to render images have become the workhorses powering world-changing AI research and development. Many of the linear algebra routines that are necessary to make Fortnite run at 120 frames per second are now powering the neural networks at the heart of cutting-edge applications of computer vision, automated speech recognition, and natural language processing.  

Now, the trend toward microchip specialization is turning into an arms race. Gartner projects that specialized chip sales for AI will double to around the US $8 billion in 2019 and reach more than $34 billion by 2023. Nvidia’s internal projections place the market for data centre GPUs (which are almost solely used to power deep learning) at $50 billion in the same time frame. In the next five years, we’ll see massive investments in custom silicon come to fruition from Amazon, ARM, Apple, IBM, Intel, Google, Microsoft, Nvidia, Qualcomm. There is also a slew of startups in the mix. CrunchBase estimates that AI chip companies, including Cerebras, Graphcore, Groq, Mythic AI, SambaNova Systems, and Wave Computing, have collectively raised more than $1 billion. 

To be clear, specialized AI chips are both important and welcomed, as they’re catalysts for transforming cutting-edge AI research into real-world applications. However, the flood of new AI chips, each one faster and more specialized than the next, will also seem like a throwback to the rise of enterprise software. We can expect cut-throat sales deals and software specialization aimed at locking developers into working with just one vendor. 

Imagine if, 15 years ago, the cloud services AWS, Azure, Box, Dropbox, and GCP all came to market within 12 to 18 months. Their mission would have been to lock in as many businesses as possible—because once you’re on one platform, it’s hard to switch to another. This type of end-user gold rush is about to happen in AI, with tens of billions of dollars, and priceless research, at stake. 

Chipmakers won’t be short on promises, and the benefits will be real. But it’s important for AI developers to understand that new chips that require new architectures could make their products slower to market—even with faster performance. In most cases, AI models are not going to be portable between different chip makers. Developers are well aware of the vendor lock-in risk posed by adopting higher-level cloud APIs, but in the past, the actual compute substrate has been standardized and homogeneous. This situation is going to change dramatically in the world of AI development.

It's quite likely that more than half of the chip industry’s revenue will soon be driven by AI and deep learning applications. Just as software begets more software, AI begets more AI. We’ve seen it many times: Companies initially focus on one problem, but ultimately solve many. For example, major automakers are striving to bring autonomous cars to the road, and their cutting-edge work in deep learning and computer vision is already having a cascading effect; the research is leading to such offshoot projects as Ford’s delivery robots.

As specialized AI chips come to market, the current chip giants and major cloud companies will probably strike exclusive deals or acquire top performing startups. This trend will fragment the AI market rather than unifying it. All that AI developers can do now is understand what’s about to happen and plan how they’ll weigh the benefits of a faster chip with the costs of building on new architectures.

Evan Sparks is CEO of Determined AI. He holds a PhD in computer science from the University of California, Berkeley, where his research focused on distributed systems for data analysis and machine learning.

How to Print an Electric Motor

How to Print an Electric Motor

Trending Technologies   August 26, 2018

An axial flux motor uses printed-circuit-board traces for electromagnetic coils

by Carl Bugeja

I started out by just wanting to make a very small drone. But I quickly realized that there was a limiting factor in just how small and light I could make any design: the motors. Even small motors are still discrete packages that have to be attached to all the other electronic and structural elements. So I began wondering if there was a way to merge these elements and save some mass.

I drew inspiration from how some radio systems used antennas made from the copper traces on a printed circuit board (PCB). Could I use something similar to create a magnetic field strong enough to drive a motor? I decided to see if I could build a motor of the axial flux type using electromagnetic coils fashioned from a PCB’s traces. In an axial flux motor, the electromagnetic coils forming the motor’s stator are mounted parallel to a disk-shaped rotor. Permanent magnets are embedded in the disk of the rotor. Driving the stator coils with alternating current causes the rotor to spin.

The first challenge was making sure I could create enough magnetic flux to turn the rotor. It’s simple enough to pattern a flat spiral coil trace and run current through it, but I limited my motor to a diameter of 16 millimeters, so that the overall motor diameter was comparable to that of the smallest off-the-shelf brushless motors. Sixteen millimeters meant I could fit only about 10 turns per spiral and 6 coils in total, arranged under the disk of the rotor. Ten turns just aren’t enough to produce a sufficient magnetic field. But the nice thing about PCBs is that it’s pretty easy today to make one with multiple layers. By printing stacks of coils, with coils on each of four layers, I was able to get 40 turns per coil, enough to turn a rotor.

This article appears in the September 2018 print issue of IEEE Spectrum as “The Printable Motor.”. To read the complete article online,  follow this link.

Self-Powered Image Sensor Could Watch You Forever

Self-Powered Image Sensor Could Watch You Forever

Trending Technologies   April 14, 2018

New technology puts the equivalent of a solar cell under each pixel

                                                                          By Samuel K. Moore

Solar cells convert light to electricity. Image sensors also convert light to electricity. If you could do them both at the same time in the same chip, you’d have the makings of a self-powered camera. Engineers at University of Michigan have recently come up with just that, an image sensor that does both things well enough to capture 15 images per second powered only by the daylight falling on it.

With such an energy harvesting imager integrated with and powering a tiny processor and wireless transceiver you could “put a small camera, almost invisible, anywhere,” says Euisik Yoon, the professor of electrical engineering and computer science at University of Michigan who led its development. They reported their results this week in IEEE Electron Device Letters.

Earlier attempts at self-powered image sensors have mostly gone one of two ways. One is to fill some of the sensor area with photovoltaics. This straightforward approach can work, but it greatly reduces the amount of light available for producing an image. The other is to have the imager’s pixels alternate between acting as a photodetector and acting as a photovoltaic cell. This too works, but at the cost of complexity and at least half the potential images.

The solution Yoon and post-doctoral researcher Sung-Yun Park came up with has neither drawback. Noting that a number of photons zip through a pixel’s photodetector diode without causing charge to accumulate, they buried a second diode beneath the photodetector to act as a photovoltaic and scoop up those strays. “It’s not really recycling; it’s more like collecting waste,” says Yoon. “It’s almost free energy.”

Because the photovoltaic is beneath the sensor, nearly all the pixel area can go to sensing the image. And because it’s using stray photons that the imaging sensor missed, it’s constantly collecting them to convert to electricity.

Though the prototype imager was constructed using standard CMOS process technology, its pixels require both a different structure and different electrical characteristics from those on a standard imager. Most obviously, the new pixel contains a p-n junction, an extra diode essentially, beneath the image sensing diode. Second, typical pixels use electrons as the main charge carrier. But to get both the photovoltaic and sensing diodes working simultaneously, Yoon and his team had to build their device so that it collects positively charged holes—electronic vacancies in the silicon—instead. Holes move less quickly than electrons in silicon, but not so slowly that it interferes with image capture.Images: University of MichiganImages from a Univeristy of Michigan’s self-powered sensor were captured at 7.5 frames per second [left] and 15 frames per second [right].

The resulting chip, with its 5 micrometer-wide pixels, was capable of the highest power harvesting density (998 picowatts per lux per square millimeter) of any energy harvesting image sensor yet. On a sunny, 60,000-lux day that’s enough power for 15 frames per second. Normal daylight conditions (20,000-30,000 lux) reduce that to 7.5 frames per second. Thirty frames per second is considered video rate, but that’s not always necessary.

Concerned only with getting a proof-of-concept chip, “we didn’t optimize the power consumption of the sensor itself,” says Park. So there is definitely room to improve the frame rate or reduce the lighting conditions needed toward what’s typical indoors. Yoon and Park have plenty of experience at that, having developed many ultralow power technologies for image sensors such as circuits that automatically adapt the frame rate to the available illumination and microwatt-scale feature detection systems.

If the project continues, they’ll work to integrate everything needed for a self-powered wireless cameras.

Content Credits: IEEE Spectrum