Hailo Unveils 26TOPS Edge AI Accelerator
Edge Artificial Intelligence (AI) technology is revolutionizing data processing within devices, facilitating swift responses in applications such as autonomous vehicles and robotics. This breakthrough innovation relies on real-time decision-making, ensuring optimal task execution in an efficient and rapid manner.
One of the significant challenges encountered by Edge AI enterprises is the development of scalable, cost-effective AI solutions that seamlessly integrate into a wide array of applications. Addressing this challenge, Hailo, an Israeli AI chip startup, has recently expanded its product portfolio with the introduction of the Hailo-8 AI accelerator series. According to Hailo, the Hailo-8 edge AI accelerator is tailor-made for entry-level applications such as video analysis, surveillance, and fundamental robot control, delivering impressive processing speeds of up to 26 tera-operations per second (TOPS).
In a recent interview with EE Times Europe, Hailo's CEO, Orr Danon, highlighted the pivotal role played by the innovative data flow architecture of the Hailo-8 accelerator in achieving cost-effectiveness and superior performance. "Rather than employing a sequential processing approach, we distribute computational tasks across various sections of the chip, enabling the simultaneous execution of different segments of neural networks. This distributed methodology harnesses both temporal and spatial dimensions, resulting in heightened parallelism and shorter data trace lengths. Consequently, remarkable performance is achieved with minimal power consumption."
Hailo's approach allows for the closer integration of different segments of neural networks (integral components of Edge AI) on the chip itself, thereby reducing the need for data to traverse long distances. This reduction in data transfer distances significantly diminishes latency while conserving energy, ultimately boosting energy efficiency, as explained by Danon. "This design empowers us to extract superior performance from every square unit of silicon, translating into substantial cost advantages. While this architecture may not be suitable for graphics processing, it excels in handling the increasingly dominant AI workloads. Thanks to optimization for parallelism, our observed data underscores the impressive power efficiency and cost-effectiveness of our approach, with the data flow architecture forming its core foundation."
Hailo has further broadened its product offering by extending the Hailo-8 Century series PCI cards, which seamlessly integrate into computer motherboards to provide additional processing capabilities, delivering processing capacities ranging from 52 to 208 TOPS, catering to applications such as video management.
Hailo asserts that when the Hailo-8 accelerator is deployed for running the ResNet 50 benchmark model (a deep neural network architecture extensively used for image classification tasks), it outperforms comparable products from Nvidia in terms of performance, cost-efficiency, and overall effectiveness.
Notably, the Hailo-8L variant runs the ResNet 50 model at an impressive speed of 500 frames per second (fps), while the Century series, designed with high-performance applications in mind, operates at a staggering 10,000 fps. The company emphasizes that, thanks to the optimized data flow architecture, Hailo-8, Hailo-8L, and Century products are exceptionally well-equipped to handle complex pipelines featuring multiple real-time streams and the concurrent processing of diverse AI tasks, offering a multifaceted solution for real-time applications.
Hailo extends its commitment to excellence by providing a comprehensive software suite encompassing a wide range of tutorials, examples, and an extensive selection of models, affording customers the freedom to customize products to suit their specific applications. It is imperative to note that the software is open source, ensuring that customers are not tethered to a single vendor and can enjoy flexibility in their choices.
