Hailo Adds Acceleration Module to its AI Edge Processor
Hailo, the up-and-coming AI chip startup, released an acceleration module this week that allows its Hailo-8 processor to be plugged into edge devices. The Israeli chipmaker is also targeting its AI modules at rivals Google and Intel, claiming its acceleration modules outperform both in neural network benchmark testing.
The M.2 and Mini PCIe AI acceleration modules are the latest in a growing roster of devices providing datacenter-level computing for edge applications ranging from industrial to consumer use cases.
Among the implementations targeted by Hailo are fan-less AI edge boxes in which multiple cameras or sensors can be connected to a single processor. The Tel Aviv-based startup said it Hailo-8 processor could be plugged into existing edge devices equipped with M.2 and Mini PCIe sockets.
The startup’s deep learning processor pegs out at 26 TOPS, with a power efficiency rating of 3 TOPS/W. Integrated with the new AI inferencing acceleration modules, Hailo claims a 13-fold higher frame rate, on average, than Google’s Edge Tensor processing unit and 26 times faster than Intel’s Myriad-X modules. Those performance results are based on multiple standard neural network benchmarks and the latest published results, Hailo said Wednesday (Sept. 30).
The benchmarks include ResNet-50, MobileNetV2, MobileNet object detection and EfficientNet-Edge-TPU classification models.
The company also claims its Hailo-8 outperforms different versions of Nvidia’s Jetson Xavier and Nano processors in ResNet, MobileNet and Yolo-V3-Tiny benchmark tests.
The AI chip startup touts its acceleration modules as smoothly integrating with machine learning frameworks like TensorFlow and the Open Neural Network Exchange, which are both supported by its Dataflow compiler. That combination would allow users to port neural network models to the Hailo-8 edge processor.
The M.2 module incorporates a PCIe 3.0, four-lane interface used to connect the AI processor to edge devices. The AI accelerator supports Linux, with Windows support coming soon, the company said.
In May, Hailo announced a partnership with manufacturing giant Foxconn Technology and a Japanese systems designer to develop the next generation of Foxconn's BOXiedge system.
The BOXiedge platform integrates the Hailo-8 deep learning processor with a parallel processor from Japanese system-on-chip designer Socienext Inc. The local video management server is designed to shift workload processing from the cloud to the edge.
With the Hailo M.2 AI module plugged into an existing BOXiedge platform, the partners said the implementation can process 20 camera streams simultaneously.
Hailo is also aiming the AI module at the automotive, smart home-city-retail segments as well as Industry 4.0 applications.