AI SoC to run Phi, Gemma, LLaVA- OneVision and Llama at edge without cloud

Ambarella unveiled N1-655 edge GenAI system-on-chip (SoC), which provides on-chip decode of 12x simultaneous 1080p30 video streams, while concurrently processing that video and running a hybrid of multiple, multimodal vision-language models (VLMs) and traditional CNNs. This SoC’s high AI processing performance supports most of the popular multimodal VLMs and large-language models (LLMs), while consuming only 20 watts of power—10-100x lower than cloud processors. For example, the N1-655 reliably runs the Phi, Gemma, LLaVA-OneVision and Llama models, without the need for an internet connection, on data inputs like visuals and speech in applications such as on-premise AI boxes, autonomous mobile robots (AMRs), and smart-city security video recorders. Following its initial N1 SoC introduced last year, Ambarella is building up a family of edge GenAI SoCs for tasks that go beyond what can be done on-camera. Both of the current family members process GenAI models locally to improve privacy with state-of-the-art performance per watt, significantly reducing power consumption as well as the total cost of ownership compared to cloud-based inference processing. “The N1-655 continues Ambarella’s industry-leading levels of AI performance per watt, this time for AMRs and on-premise AI box applications that require the latest edge GenAI features while keeping power consumption under ...