Issue link: https://iconnect007.uberflip.com/i/1345210
42 SMT007 MAGAZINE I MARCH 2021 Companies are moving to make edge devices smarter in automo- tive, drones, security, etc. at is where we are focused. Deep learning is everywhere—you can use deep learning in computer vision, audio, communication, and more. Hailo is focused on vision-based applications. Obviously, there are many, many types of applications that you can see here. Also, we are seeing an acceleration of additional applica- tions in smart retail, medical robots, last-mile delivery, and industrial robotics, which can likely be attributed to impacts from COVID-19. Johnson: Yes, there has been a shi in industrial robots, that is for sure. Bar: at's relevant for our discussion on Indus- try 4.0. It's worth noting here that we raised our Series B round just last year, bringing our total funding to $88 million. And in the last round, we had two new strategic investors: NEC and ABB, the latter a major player in Industry 4.0 and industrial robotics. It's important for me to give you more con- text on our processor as well. e Hailo-8 was designed internally from the ground up. It's a novel technology that we developed. ere is no CPU/GPU inside. It includes 26 TOPS (Tera Operation per Second) in 8-bits, giving it high-power efficiency. at's quite a lot if you compare it to other available solutions. e innovation comes from the Hailo-8 archi- tecture, and one noticeable and important fac- tor is the fact that the Hailo-8 does not require external memory. When you are talking about running neural networks, this is very important, because in traditional architecture, every time you transfer data from one layer to another, you need to access the DDR. As a result, you will add power and latency. In some cases, specifically on the industrial side when you have a pick-and- place machine, latency is very important. In terms of the connectiv- ity, Hailo's chip can be used as a standalone or as a co-processor. As of today, most of our engage- ments are using the Hailo-8 as a co-processor. Customers already have their product deployed in the field, but they are look- ing to improve the performance by overcoming the limitations involved with running neural net- works at the edge. For those cus- tomers, we offer them a very fast time-to-mar- ket solution by connecting the Hailo-8 through high-speed interfaces such as PCIe and ETH, and to offload all the tasks related to neural networks to the Hailo-8. Johnson: It's logical, at least to me, that more computation and analysis needs to happen at the edge within the sensor device itself. en you're communicating data that is already par- tially analyzed and that frees up the central com- puter resources to work on higher-level things. Bar: Definitely. e efficiency is very impor- tant. For example, some companies in this space offer solutions connected to the host accelerator, which is a fixed HW pipeline. You don't necessarily have better performance when you have more compute elements. Why? Because you will have a bottleneck some- where in between the accelerator and the host; the AI HW accelerator finished the work very quickly, but the host cannot process it with the same speed as the HW accelerator, so it's meaningless. You need to think outside the box and design a dedicated processor such as the Hailo-8 to run DL-based applications. Hailo's innovative structure-defined data flow has a completely redesigned processor architecture, with a distributed on-chip mem- ory fabric (no external memory), novel con- trol schemes, efficient interconnect, and a full-stack soware toolchain called "data flow compiler." Liran Bar