Hailo ports transformer AI to embedded chip – updated
Israeli AI chip startup Hailo has ported an AI framework using transformers to its architecture at the request of a customer.
Transformer networks are very different from the embedded AI such as TinyML used in the Internet of Things or the image recognition AI frameworks such as YOLO or ResNet. Running transformers locally in embedded chip can reduce the overall power consumption and provide more responsiveness for transcription and generating language.
The transformer AI framework was run through Hailo’s compiler at the request of a customer, Eyal Barnea, vice president of product at Hailo, tells eeNews Europe. This is running on the first generation Hailo-8 chip and handling six HD video streams.
Hailo launched its second generation Hailo-15 family of high-performance vision AI processors in March for integration directly into intelligent cameras.
- Second generation AI vision processor targets smart cameras
- Hailo, Renesas team up for seamless transition from ADAS
“Hailo-15 represents a significant step forward in making AI at the edge more scalable and affordable,” said Orr Danon, CEO of Hailo. “We are leveraging our leadership in edge solutions, which are already deployed by hundreds of customers worldwide; the maturity of our AI technology; and our comprehensive software suite, to enable high performance AI in a camera form-factor.”
The Hailo-15 family has three variants, the Hailo-15H, Hailo-15M, and Hailo-15L, ranging from 7 TOPS (Tera Operation per Second) to 20 TOPS. They all support multiple input streams at 4K resolution and combine a CPU and DSP subsystems with Hailo’s field-proven AI core based on an array of multiply accumulate (MAC) blocks.
This gives the flexibility to support new types of AI framework such as transformers and LLMs. Hailo is using vision-based transformers in cameras for real-time object detection. The added AI capacity in the Hailo 15 can also be used for video enhancement and much better video quality in low-light environments, for video stabilization, and high dynamic range performance.
As an example, the Hailo-15H is capable of running the state-of-the-art object detection model YoloV5M6 with high input resolution (1280×1280) at real time sensor rate, or the industry classification model benchmark, ResNet-50, at 700 frame/s
If you enjoyed this article, you will like the following ones: don't miss them by subscribing to :
