Will Analog Chip Technology Become Relevant for AI?


A Dialogue with Kurt Busch of Syntiant

Due to the widespread applicability of the nature of AI, one can use it in a wide range of applications that need a wide range of power and performance. That ultimately boils down to the need for chipsets that offer required power and performance for a given application. This has caught the attention of large and small companies, as well as startups.

While digital technology is the most prominent fundamental technology for AI chipsets, alternative approaches, such as neuromorphic and analog, have been proposed. One AI chipset startup, Syntiant, based in Southern California, is using analog technology to solve the AI acceleration problem and is promising to reduce power consumption radically in comparison with the digital technology. The company recently secured funding from Intel Capital and Tractica recently had an opportunity to speak with the CEO, Kurt Busch. Here are some of the excerpts from our discussion.

What are the key aspects of your technology that differentiate you?

We are focusing on analog compute in memory technology, unlike the digital stored program architectures used in the mainstream chipsets today. Our analog in-memory computation eliminates the largest consumer of power in deep learning (DL) inference, which is the penalty paid by moving data in and out of memory. We are also using singled transistors to perform an analog multiply and accumulate (MAC) operation, which is much more power and space efficient than a digital MAC. This promises to reduce the power consumption by a factor of at least 50X (in terms of operations per watt) versus current machine learning (ML)-optimized solutions. Our analog technology brings down the overall chip power consumption of milliwatts into microwatts, paving the way for AI technology to be used in a wide range of applications where available power is limited.

What do you see as some of the key technological challenges in productizing and mass production of your technology?

Analog neural networks have been academically interesting for decades, but without commercial deployment. The primary challenge with any analog technology is making it mass producible and having a practical programming model. Syntiant is dedicating a large amount of effort engineering a solution that can ship in the billions of units and has focused on common ML frameworks to make programming transparent.

How does your software development flow work? How do you map digital parameters to analog and store on the chip?

Programmability is one of the key advantages of Syntiant technology. The neural networks are trained using traditional frameworks, such as TensorFlow, in a manner indistinguishable from targeting idealized inference deployment. The resultant neural network parameters are programmed directly in the chip as firmware with Syntiant-proprietary algorithms, thus making it easy for developers.

What is your monetization strategy?

We are a semiconductor company, so we sell chips. We plan to support three primary models:

  1. We support customers with no ML capabilities. Here, we will supply an out-of-the-box solution, complete with training services to enable practically any device.
  2. For companies with strong ML capabilities, we will share our software flow, so they can train our devices for whatever application they would like.
  3. For large customers that have an ML solution they have qualified in the cloud and would like to have a low-power custom application specific integrated circuit (ASIC), we can support building custom devices using the same tools we developed to support our own.

Which use cases are you targeting?

Our low-power technology is a natural fit for battery-operated devices. Our technology promises to increase the battery life from days to months and we plan to target the always-on inference market segment. We are focusing on DL technology and our initial set of customers is targeting classification problems, such as voice recognition, speaker ID, and image recondition, as well as sensor synthesis.

And finally, congratulations on your funding from Intel Capital. How will funds help you grow your business?

We are very excited to be working with Intel Capital and these funds will allow us to sample our first silicon with customers later this year.

Comments are closed.