Expedera: Custom Deep Learning Accelerators Through Soft-IP

Configuring deep learning accelerator hardware through software.


Internet of Things (IoT) and Artificial Intelligence (AI) have caused a massive increase in data generation — and along with it, a need to process data faster and more efficiently.

Dubbed a “tsunami of data,” data centers are expected to consume about one-fifth of worldwide energy before 2030. This data explosion is driving a wave of startups looking to gain a foothold in custom accelerators.

With a projected compound annual growth rate (CAGR) of 39% from 2021 to 2027, the global market for data center accelerators is predicted to reach $69.4 million in the next five years. The increasing demand for deep learning – a subset of AI – is an important element in this growth. This is where startup Expedera is making its move.

While custom deep learning accelerators (DLAs) are nothing new, Expedera is taking a different approach — semiconductor IP. Instead of optimizing power, performance and cost-efficiency through the typical route of chips, the startup is using soft IP to configure neural networks.

“Where we really differ is the fact that our IP is made up of fundamental building blocks that allow us to do almost an infinite amount of calibration for what a customer is looking for,” said Paul Karazuba, Expedera’s vice president of marketing.

The company claims to be entirely technology agnostic. Its product is “just a configuration that you put into your hardware design,” according to Siyad Ma, co-founder and vice president of engineering. “You start with a configurable hardware where you can specify how much compute power you need, how much memory you need, etc., and that will just produce the hardware that you want. In a software tool chain, there’s also some programmability that allows you to specify what you want.”

Hardware is almost always faster, but software configurations are improving. More designers of machine learning systems are demanding software development kits (SDKs), and solutions that heavily weight hardware versus software are more difficult to work with and much more difficult to adapt and update.

“All you’re doing is executing layers and layers of neural network definitions, and if you can manage to make each layer run efficiently, you’re not compromising the overall execution,” Ma said.

Co-founded by Ma and Da Chuang, Expedera designed its DLAs specifically for AI inference. The applications for inference are “theoretically limitless.” This can include low light noise reduction in 4k streaming video, a use case that went into production last month for the company’s first customer, or something as basic as audio background noise reduction in a car.

Rather than having to choose from “cut A or cut B,” Expedera is emphasizing “more degrees of freedom,” allowing its customers to develop their ideal neural network engines, Karazuba said. “The use cases of AI are limitless — and that’s honestly one of the biggest problems with AI right now. There are so many things that could be done with it that the customers don’t know what they actually want to do.”

Ma noted the company expects to see two types of customers — those with specific applications they want to run, and those that just want to open a door and be able to run numerous future applications.

Expedera is targeting automotive, data centers, edge, and smartphones as its markets, as AI inference is rapidly dominating those industries.

From a security perspective, the company said its implementations are not any more or less secure than any other implementations might be, but because its product is not widely known in the market, security will be less of an issue, at least initially.

The company emerged from stealth in April 2021 and secured $18 million in a Series A funding round led by Dr. Sehat Sutardja and Weili Dai, the founders of Marvell Technology Group. Expedera’s total investment is currently at $27 million.

Leave a Reply

(Note: This name will be displayed publicly)