Will Analog Signal Processing be the Future of AI?
Article By : Nitin Dahad
Gene Frantz may have been the visionary for digital signal processing (DSP) back in the 1970s, but now he thinks we need to turn our attention back to analog to tackle the big challenges of artificial intelligence
Previously a principal technology fellow at Texas Instruments, Frantz is now a professor at Rice University. He is also the co-founder and chief technology officer at at Octavo Systems, a fledgling semiconductor-in-package (SiP) company based in Austin, Texas.
Speaking during the launch of Octavo’s OSD32MP1 — the company's first SiP based on the newly announced STMicroelectronics STM32MP1 microprocessor — Frantz told EE Times that he believes SiP and analog processing will be the future. He said AI needs a better solution and suggested that we should consider going back to analog signal processing.
“When most people listen to the words ‘analog signal processing’ they probably think analog computing, but that’s not really what I am saying," Frantz said. "If I can take the whole idea of signal processing and do an analog arithmetic logic unit (ALU) or mixed signal ALU, I can increase the performance by orders of magnitude, and at the same time reduce the power dissipation by orders of magnitude. And the only problem with that is that I have an issue with dynamic range, with accuracy and with linearity. Those are major issues. But the question is, if I can give you three or four orders of magnitude of higher performance, and three or four orders of magnitude lower power dissipation at the same time, do you think those three problems can be solved?”
Speaking of the challenge, Frantz used the analogy of a junior high school dance, where the boys line up on one side and the girls line up on the other, but nobody comes to the center to dance. “I’m watching all the people who know the theory of signal processing, and all the architects of processing systems, standing on their walls and not coming in to dance," he said. "And it’s going to take a combination…just like a junior high school dance, a lot of the activity actually goes on in the hallways, not in the dance hall.”
Frantz at the Octavo booth at Embedded World. (Nitin Dahad/EE Times)
Frantz said SiPs are beginning to change that. “Now I can do exactly what I want to do because I can’t put a good analog signal processing capability in the same process as a good digital signal processing capability, but I can put them next to each other, which is what we are doing in the SiP,” he said.
SiP Vision: No Pins
Frantz said with shrinking device sizes his end goal with SiP technology is to someday have a single package that generates its own energy or creates its own energy, has its sensor base, has its control or computer and communicates wirelessly, and has no pins.
Years ago, Frantz said he and Masood Murtaza — who led packaging at TI along with Frantz — began talking about Moore's Law and its measure of success in process silicon. "What is the measure of success in system in package? The answer to that is fewer and fewer pins," he said. "And we felt the ultimate single system-in-package had zero pins.”
Extending Moore's Law
Tongue-in-cheek, Frantz said his dream is to have walls painted with SiPs, enabling anyone to change the color of the walls with a remote control. "The reason I’d like to do that is, if you can add a poster child to a concept, people may laugh at your poster child, but they’ll understand the concept,” he said.
This isn't really a new idea on Frantz's part, he pointed out, adding that the University of California-Berkeley came up with the idea of smart dust about 20 years ago. “All I’m doing is saying, as technology advances, we are getting more and more capable of actually doing that,” he said.
“If you think I have all these parts and they are connected by RF interconnects and I network them, not only can I make it a solid color, but I can make it different colors. I might even be able to do art with it," Frantz said. "I spent a good part of my career looking at things that were considered impossible."
In the next decade, half of the semiconductor market will be driven by SiP to address the need for more intelligence at the end nodes, according to Frantz. “What we are doing is taking the philosophy that SoC isn’t really a system on chip, it’s a sub-system on chip, and it always goes into a bigger system,” he said.
Octavo's OSD32MP1 SiP product is based on STMicroelectronics STM32MP1 microprocessor and complements Octavo's family of modules based on the Texas Instruments AM335X processor. (Source: Octavo)
“And we are making it more and more possible for the semiconductor process to tolerate lower volume per product, which system integration requires. If I do transistors, I can do hundreds of billions of them and everyone uses them, and if I am doing a specific system, not many people can use it. As we go that way, you begin to see that the only way I can do integration is at the system level, using something like SiP. And it’s going to be lower volume, which is going to take over the market.”
Frantz added that the whole area of AI and deep learning will be better addressed once the relationship between performance, power dissipation and size will be understood to the point that people can do on their smartphones in real-time the magnitude of computing that occurs only in the cloud today.
Frantz said that SiP technology can also play a role in extending Moore's Law, at least in a way. “When you are moving from system-on-chip to SiP, you’re putting more transistors on the same substrate, so in fact we are still continuing to achieve Moore’s Law," he said. "It’s just that innovation is moving from the silicon process to the assembly process.”
Practical approaches to driving more computing power in the future — including moving to heterogeneous integration — have been gaining steam. Semiconductor-in-package is one form of this, and Frantz’s idea of creating a wall consisting of an array of RF interconnected SiPs that can then be programmed to configure the art as you please, is another.