NXP intends to purchase AI-chip startup Kinara, closing the deal by midyear. Kinara’s NPUs complement NXP’s i.MX line of low-cost Arm microprocessors and i.MX RT “crossover microcontrollers” (even lower-cost microprocessors). We expect future NXP processors to integrate Kinara’s technology.
Targeting automotive and industrial markets with its microprocessors and microcontrollers, NXP was early to integrate AI hardware into “edge” (embedded) chips. The company licensed accelerators before developing its Neutron NPU. Nominally scalable to multiple TOPS and capable of executing transformer-based neural networks such as large language models (LLMs), Neutron first found use in the recent MX RT700 and i.MX 94 devices and delivered much less than 1 TOPS.
By contrast, Kinara’s original Ara-1 device’s maximum throughput is 6 TOPS, and the subsequent Ara-2 peaks at 40 TOPS. Moreover, the company has demonstrated Ara-2 running the Llama-7B LLM and generating 12 output tokens per second while requiring only 2 W (typical). Thus, Kinara provides NXP a high-throughput, transformer-capable NPU, two things NXP must have to be relevant in edge AI.
Like other edge-AI products, Ara’s success, however, has been in vision applications. When it entered production in 2021, the Ara-1 stood out for its high throughput on CNN-based models and modest power. Its architecture achieved much greater actual performance than contemporaneous NPUs with similar TOPS ratings because it could better utilize its math engines. Moreover, Kinara optimized the design to perform well on single images instead of requiring latency-inducing batching to achieve high throughput.
Looking Forward
We expect NXP to merge Kinara’s stack with its eIQ software and integrate the company’s NPU with i.MX and i.MX RT chips, reducing the number of neural cores to scale down performance and power to the i.MX RT’s proportions. It would also make sense for NXP to develop a safety-certifiable version for automotive applications. Many companies are pursuing low-cost, low-power NPUs. Through this acquisition, NXP gains an efficient architecture proven in customer designs, and Kinara doesn’t need to worry about funding its next chip.