NXP Semiconductors has acquired Silicon Valley edge AI chip startup Kinara for $307 million in cash. The AI chip startup raised $54 million in funding, completing a B round in 2021. The deal follows recent exits by edge AI chip competitors GrAI Matter (acquired by Snap), Perceive (acquired by Amazon), Flex Logix (acquired by ADI) and Blaize (floated via SPAC).

Kinara’s technology and markets fit perfectly with NXP’s AI strategy, Rutger Vrijen, senior VP of strategy at NXP, told EE Times.

“We believe that the ability to run AI at the edge has massive advantages over having to rely on the cloud,” Vrijen said. “This comes down to energy efficiency and cost, but also the reduction of latency allowing devices and equipment to make real-time decisions.”

Part of the vision is human-like interactions in real time with edge devices, which would require multi-modal transformer networks running on device, an area Kinara specializes in.

“By bringing Kinara into our fold, we’re redefining what’s possible at the edge with new classes of powered systems,” Vrijen said. “By new, we mean new combinations of their discrete NPUs and our processing portfolio.”

Kinara has an existing partnership with NXP, whereby the companies had demonstrated Kinara AI accelerators offloading AI workloads from NXP application processors. A recent demo at CES showed LlaVa-7B performing image classification plus Llama, which was used to describe images in natural language, both running on Kinara with an NXP host. An example use case might be a home security camera system that describes for the user what happened while they were out.

“We could see the enthusiasm that it created in our customers that we showed this to,” Vrijen said.

Multi-modal transformers

Kinara had developed two generations of its AI accelerator. The first-generation Ara-1 is vision focused and aimed at smart retail and smart city applications, while the second-generation Ara-2 is aimed at LLMs and multi-modal transformers with up to 30 billion (INT4) parameters.

“What we found really appealing about adding a focused NPU solution to our portfolio is the ability to scale easily,” Vrijen said. “We can pair with one Ara-2 or multiple Ara-2s to rapidly scale up from Ara-2’s 40 TOPS in multiples of 40, for use cases that require it. We couldn’t do that if we had added a 40 TOPS NPU inside an SoC.”

Kinara IP will not replace NXP’s existing home-grown Neutron AI accelerator IP, which features on some NXP chips already.

“We have a strong roadmap for [Neutron],” Vrijen said. “Clearly, we’ll explore future options to leverage the strength of both IP blocks, but we haven’t made any decisions yet on when and how to integrate Kinara IP into our MPU portfolio at this point.”

NXP will continue to offer both Ara-1 and Ara-2 as discrete chips and will support Kinara’s ongoing roadmap.

Kinara Ara-2

“The immediate plan is to support Kinara’s existing products as well as their own NPU roadmap and match it up with our processor portfolio in a discrete fashion…we’ll take our time to learn from customers and the use cases they deploy to see what might drive a tighter integration into a SoC, if at all.”

Vrijen added that future integration could also be in the form of co-packaged die or modules, before potentially moving to a full SoC.

Both Ara-1 and Ara-2 can be used with NXP’s i.MX-RT crossover processors or i.MX application processors as a host, but new and existing Kinara customers will be served by NXP—no matter which host processor they are using, Vrijen said.

“We believe our portfolio is stronger together,” he said. “The more we can add by providing our customers with pre-tested pre-integrated solutions, the more they will benefit from that, from a go-to-market perspective and a development time perspective, but we’re not going to prohibit people from matching Kinara’s products with other hosts if they choose to do so—you can always buy our SKUs individually, if you want.”

Kinara has sold about 500,000 of its AI chips up to now, Vrijen said, noting these volumes are mostly for pilot programs and use case evaluations. Kinara’s customers are mostly in the industrial and consumer sectors, but NXP wants to market Kinara chips to its automotive customers, too. Vrijen said Kinara chips would be offered for automotive infotainment applications—perhaps something like a PC copilot in the car—or in-cabin monitoring, rather than for ADAS or autonomous driving.

NXP’s ambition is to offer end-to-end solutions as far as possible, Ajith Mekkoth, senior VP of engineering in NXP’s secure connected edge group, told EE Times.

“One of the appealing things about a discrete NPU is that the AI workload is evolving,” Mekkoth said. “Until that settles down, it may not be prudent to jump into integration discussions. We’ll need to understand where GenAI workloads will plateau, and then look at integrating the right kind of compute structure, and that’s something we’ll watch closely. We’re hoping customers will find the discrete solution quite flexible in the meantime.”

Mekkoth said Kinara’s mature software toolchain was also a big part of its appeal.

“A big threshold to adoption is the ability to actually use the silicon,” he said. “Kinara understands that in order to make their silicon useful, [customers] need to be able to easily compress models into these chips, and compile models they may have developed themselves, and that’s where they have put a lot of their energy.”

It is likely Kinara’s toolchain will be incorporated into NXP’s eIQ ecosystem going forward, he added.

Kinara has around 100 people, mostly engineers, working across Silicon Valley and Hyderabad, India. Vrijen expects NXP will fully absorb Kinara and all of its U.S. and India team.

The deal is still subject to regulatory approval but is expected to close in the first half of 2025.

By: DocMemory
Copyright © 2023 CST, Inc. All Rights Reserved