Demand is growing for AI solutions that process images from cameras and other photo sensors. These solutions have applications in security, autonomous vehicles, healthcare, smart cities, and many other areas, but the more advanced they become, the more computationally intensive they get. So creating more efficient AI computational devices is a necessity for advancing the field.
Chip maker Qualcomm held a series of briefings late last month that point to a solution on this front. The company discussed an R&D project it has undertaken to reduce the amount of compute necessary to do visual AI, and thus create chips that are smaller, more cost effective, and use less power.
The techniques it is implementing include creating methods to eliminate redundant information processing by only analyzing frame differences rather than analyzing each frame. In typical videos, each subsequent frame contains little information change from the previous one, so not having to process each frame individually with duplicated information is much more efficient. Further, Qualcomm is developing a skip function that limits the number of frames to be analyzed by skipping frames that offer little or no change, eliminating the computational burden of processing them. This creates a process flow that understands the relationship between frames enabling the system to exit processing of the visual data earlier when no additional information is needed, and potentially saving many frame processing cycles.
The goal to reduce the complexity of the processor so that the amount of power used by the chip can be significantly reduced and the actual size of the chip created to be more compact. Both of these capabilities have the potential to reduce the cost of chips. Smaller less power-hungry chips produce less heat and also enable smaller finished devices with more limited power needs. That means being able to move the chip right into the camera or similar products without needing an external processing component as is common in current systems.
A further benefit of this work will be reducing the need for complex processing in the cloud or at the edge. The less data sent to the cloud for processing, the less cost involved in transport. This results in faster data analysis with less latency, less sharing of the data for increased privacy/security, and a reduced cloud processing load.
Indeed, edge computing is becoming commonplace in distributed, often hybrid cloud-based environments. As the number of cameras and visual devices proliferate, the workloads placed on edge computing systems increases, making the deployment more complex and more expensive. AI programs that can be embedded in visual devices would enable a large scale deployment of smart cameras for security, smart cities, autonomous vehicles, etc. while reducing overall system costs. Complexity and deployment cost are a prime inhibitor to greater use of such solutions in both public and private markets.
Of course, Qualcomm isn’t the only company working to find a solution to this challenge. Intel/Movidius and NVidia are key players in AI-based video systems that have current specialized offerings on the market that also enable the ability to do embedded visual processing. And other major players are experimenting in this space, including Google, Microsoft, and several players that supply mobile systems (e.g., Samsung, NXP, ARM, etc.). Each has implemented its own designs based on its unique algorithm acceleration, but more generic devices (e.g., Nvidia) may not be as effective at visual processing tasks. Qualcomm also has the advantage of being a supplier of low power processing solutions inherited from its mobile processing strengths, so it has a strong advantage when it comes to impacting embedded solutions at the edge. But this is a very competitive market that will grow over the next several years and will likely not be dominated by any single vendor for the foreseeable future.
Bottom line
By creating a way to reduce the processing required to do visual AI computations and thereby reducing the power requirements, size, and cost of AI chips, Qualcomm hopes to create mass market devices that can be deployed in lower cost and higher quantity devices, and without sacrificing AI quality. If it can accomplish this, the number of devices with self contained capability will skyrocket, relieving the strain on edge computing and allowing it to do more important processing functions. This is a win for Qualcomm, and potentially a win for all forms of new, visually “smart” products. While Qualcomm is not the only silicon vendor in this race, its R&D efforts should put it ahead of many, and it should benefit significantly from its leadership.
Jack Gold is the founder and principal analyst at J.Gold Associates, LLC., an information technology analyst firm based in Northborough, MA., covering the many aspects of business and consumer computing and emerging technologies. Follow him on Twitter @jckgld or LinkedIn at https://www.linkedin.com/in/jckgld.
