Skip to content
Israeli chip startup Hailo has launched three new visual AI chips designed for different use cases, unlike the Hailo-8. While the Hailo-8 has primarily been used as a vision processing co-processor in industrial edge boxes and aggregators, the Hailo-15 series is a SoC with an accelerator, CPU, and DSP aimed at IP cameras.
Hailo will continue to produce the Hailo-8 and introduce a new product line.
Hailo CEO Orr Danon said, “Many customers sometimes do AI processing on the camera and sometimes on the gateway, ideally wanting to reuse software investments across different applications. We are launching a new product line for smart cameras that relies on the same technology stack as the Hailo-8… This stack is new and improved but fully compatible and can be deployed in various forms.”
Although customers typically purchase either the Hailo-8 or Hailo-15, both may be used simultaneously in some setups, Danon said, such as performing vision processing at the endpoint and then aggregating streams from multiple cameras at the gateway.
CPU and DSP
The Hailo-15 series offers heterogeneous computing, including a CPU for orchestrating tasks, DSP for flexible yet more specialized computing, and an accelerator for handling AI inference.
Danon said, “AI is not just about neural networks, whether CNN or transformer, there are multiple aspects, and DSP is an effective tool to bridge the gap between CPU and AI accelerator.” He added that both CPU and DSP are licensed from third parties and are not based on RISC-V.
Danon stated, “The processing power of the Hailo-15 is weaker than that of the Hailo-8, its AI accelerator architecture is similar to the Hailo-8 but improved.”
The Hailo-15H provides 20 TOPS, the Hailo-15M offers 11 TOPS, and the Hailo-15L delivers 7 TOPS of INT8 performance (the Hailo-8 provides 26 TOPS). He added that power consumption depends on the workload, but current customer operations run below 2W.
These performance levels will suit smart cameras with multiple sensors, depending on how many tasks they want to run simultaneously and the required resolution, frame rate, and prediction accuracy.
He said, “Our customers are running over 20 tasks simultaneously on such devices… This is heavy processing, but some customers prefer more compact, cost-effective solutions with lower performance requirements.”
Many Hailo customers will also use AI to enhance image quality, including low-light performance, digital zoom, etc. These applications are common in smartphone cameras but have yet to enter IP cameras.
Hailo-15 Supports Transformers
Hailo’s software stack already supports transformers, including ViT for the Hailo-8, and this feature will also be available for the Hailo-15 series.
Danon said, “Our customers have deployed transformers on devices. Transformers are becoming increasingly powerful, but I don’t think they will completely replace CNNs. They excel in certain tasks, but in some tasks, they are not more efficient than CNNs. We also see some very powerful combinations.”
The key advantage of transformers is adding context to what is happening in an image. Danon cited the classic example of bus advertisements. Transformers can provide context for person detection offered by CNN, helping to distinguish between the images of people in advertisements and real pedestrians.
He said another powerful combination of CNNs and transformers could be adding natural language descriptions to images.
In addition to supporting transformers, Hailo’s software stack has also expanded support for different models through the company’s model library and added training environments for these models to facilitate productization.
While early samples have already been delivered to customers, the Hailo-15 series will begin offering samples mid-year.
Danon indicated that architecture details for the Hailo-15 will be revealed later this year.
Hailo Adds Vision Processor SoCs for Smart Cameras — Sally Ward-Foxton
from A to B
