Kinara's Ara-2 Edge AI Chip Is Up to Eight Times Faster, Can Run Large Language Models on Device
Delivering a claimed 485 frames per second in ResNet50 and enough horsepower for on-device generative AI, the Ara-2 is a miniature marvel.
Kinara, a specialist in energy-efficient artificial intelligence at the edge, has unveiled its Ara-2 processor — claiming enough power to run large language models (LLMs) and other generative AI models on device, with up to eight times the performance of its predecessor.
"With Ara-2 added to our family of processors, we can better provide customers with performance and cost options to meet their requirements," claims Kinara's chief executive officer, Ravi Annavajjhala, of the new part. "For example, Ara-1 is the right solution for smart cameras as well as edge AI appliances with 2-8 video streams, whereas Ara-2 is strongly suited for handling 16-32+ video streams fed into edge servers, as well as laptops, and even high-end cameras.
"The Ara-2 enables better object detection, recognition, and tracking," Annavajjhala continues, "by using its advanced compute engines to process higher resolution images more quickly and with significantly higher accuracy. And as an example of its capabilities for processing Generative AI models, Ara-2 can hit 10 seconds per image for Stable Diffusion and tens of tokens/sec for LLaMA-7B."
The chip is designed to replace the original Ara-1 for those requiring more performance, and it's an undeniably impressive upgrade: the company claims the part can deliver between five and eight times the performance of Ara-1 — and that it's powerful enough to take the place of higher-cost and more power-hungry graphics processors for a number of models including, but not limited to, large language models (LLMs).
Kinara is due to showcase the Ara-2 at the Consumer Electronics Show (CES) in January, and has confirmed that the part will be available as a bare chip as well as in single-chip USB and M.2 modules and a PCI Express add-in board featuring four Ara-2 chips operating in parallel. Pricing, however, has not been made public.
More information is available on Kinara's website.