CEO interview: with Kurt Busch of ‘always-on’ startup Syntiant

CEO interview: with Kurt Busch of ‘always-on’ startup Syntiant

Interviews |
By Peter Clarke

Syntiant Corp. (Irvine, Calif.), founded in 2017, is one of the more successful ‘edge’ AI processor companies having shipped tens of millions of neural processors across applications in consumer and industrial electronics. eeNews Europe caught up with Kurt Busch, CEO of Syntiant, to find out more about the company’s success and its ambitions.

Busch starts by pointing out that the start of the company was about finding ways to increase the efficiency of the multiply accumulate (MAC) operations that are the majority of those used in neural network computing. One of the ways Syntiant tried was to use analog operations within flash memory. “We learned that detecting 256 levels within a flash memory cell designed for two levels is challenging. But the biggest problem is that flash is not very temperature stable. So, it was in 2019 that we eventually decided to adopt a digital methodology and an ‘at-memory’ rather than ‘in-memory’ approach,” said Busch.

So, if the density and the power consumption of non-volatile flash are not Syntiant’s AI market advantage, what is?

Busch sums it up by saying that success often comes by taking as broad a view of the task in hand as possible. And this will usually be best served by a combined software and hardware platform.

“We strive to minimize the memory movement in our Neural Decision Processors,” said Busch. And this is done partly by focusing on the machine learning models and partly by choosing the ‘edge’ applications to match the resources, he added. To this end Syntiant has focused on being the ‘always-on’ gatekeeper for systems that can be put into low-power ‘sleep’ modes – and on developing best-in-class algorithms for such things as wake-word recognition and event detection.

The company started with its NDP100 series of neural decision processors based on the Core1, mainly focused on audio streams. Then it introduced the Core2 to power the NDP200. The much-improved AI core is able to discriminate movement and person-detection in visual applications as well as performing audio inferencing.

By matching the single-chip AI resource to the application Syntiant has been successful at keeping whole models on-chip and avoiding excessive movement of data on an off chip, which is problematic for many AI accelerators trying to tackle more complex problems at the edge. But getting the models minimized yet well-tuned is key, said Busch.

Model like you mean it

“About 75 percent of our employees work on software and about 75 percent are machine learning focused,” said Busch. “And about 50 percent of our business is licensing software to companies. The rest is selling our processors, usually bundled with our software. The difference between demo machine learning models and production models is immense.”

And the Syntiant ML software is hardware agnostic Busch emphasized. It is written to run natively in compute-constained environments including Arm, x86, DSPs, GPUs, other NPUs and FPGAs and neural network ASICs. This means that established and even deployed hardware architectures can be upgraded with Syntiant edge AI software prior to a redesign that could incorporate a Syntiant NDP.

But Busch’s bottom line is: “We have the best models.” He provides an example. How to detect an external window breaking while ignoring household noise. Conventional demonstration software might recognize a specific sound profile in a key frequency band. But this can miss many glass-break events (false negatives) and be triggered by many loud noises (false positives). To improve on this Syntiant collected thousands of examples of external home glass breaking and trained advanced neural networks to trigger only when an external glass break is detected and ignore other noises, even plates broken in the kitchen.

Syntiant models are hardware agnostic. Source: Syntiant.

Avoiding those failures is the key to power saving and customer satisfaction, said Busch. That comes from really understanding the broad context, he said.

Syntiant has chosen to work in the ‘always-on’ domain and is happy to continue there.

There are a lot of AI startup players addressing a broader set of edge applications but Busch said Syntiant feels no need to break out of the niche where it is enjoying success and that has a long way to grow as a market. “Always-on is about the efficiency of the interface. We’ve gone from keyboards, to mice, to touchscreens, to speech and sensors,” said Busch.

This provides a hint of the direction Syntiant is moving in – to seeing and feeling at the edge.

Core 3 coming

But more complex applications will require more capable NDPs if the inferencing is going to be contained within a single chip or chiplet-enabled component. Both the NDP100 and the NDP200 series processors were designed for TSMC’s 40ULP 40nm ultralow power manufacturing process.

Back in March 2022 Syntiant said it expected to introduce the Syntiant Core 3 processor in 2023.

Busch confirmed that Core3 is targeting a more advance manufacturing process node than the 40nm 40ULP and said that he expects Syntiant to have silicon in house before the end of 2023. He declined to name the targeted process node or the performance uplift the company is seeking.

Core3 and machine learning algorithms it is preparing to run on it could provide another reason to seek out Syntiant at CES or other exhibitions in the remainder of 2023 and 2024.

Related links and articles:

News articles:

Blumind startup pursues analog AI at the edge

Startup Syntiant ships 20 million AI processors, raises $55 million

OKdo selected by Google sensor spin-off as SBC partner

Analog machine learning chip lowers ‘always-on’ power

If you enjoyed this article, you will like the following ones: don't miss them by subscribing to :    eeNews on Google News


Linked Articles