Real-Time Scaling for Smart City Applications
By Michael Chang
Vice President Product Marketing, NVXL Technology, Inc.
According to smartcityhub.com¹, it is estimated that by 2040, more than 65 percent of the world’s population will be living in cities. This is placing a huge burden on existing urban infrastructure and is driving innovation in smart city technologies. A smart city is an urban area that uses different types of electronic data collection sensors to supply information to manage resources efficiently. This includes data collected from citizens, devices, and assets that is processed and analyzed to monitor and manage traffic and transportation systems, power plants, water supply networks, waste management, law enforcement, information systems, schools, libraries, hospitals, and other community services. With the combination of the latest low power sensors, wireless networks, and web-and mobile based applications, smart city technology is already a reality in a number of cities around the world.

Research from IHS markit² says the number of smart cities is expected to quadruple from 2013 to 2025 worldwide. Cities around the world are expected to spend as much as $41 trillion on smart tech over the next two decades. China is at the forefront of this trend, leading the way with the most smart city projects worldwide. A Frost & Sullivan study predicts that by 2025, more than half of Asian smart cities will be in China, which has already announced that 500 of its cities would undergo smart city transformations with associated projects projected to generate $320 billion for China’s economy. As part of driving this growth, there are already more than 170 million surveillance cameras in China and the country is expected to have another 570 million cameras by 2020, that is nearly one camera for every two citizens. As the construction of smart cities accelerates, the scale of related markets, including surveillance cameras, video analytics, etc., is expected to hit 100 billion RMB (14.7B U.S.). This figure rockets up to 1 trillion RMB (14.7T U.S.) when upstream and downstream industries are included.

With the advance of deep learning algorithms, improved image recognition and object detection capability and accuracy, computer vision is gaining traction in smart city applications where objects like cars, pedestrians or motorcycles are detected and analyzed to optimize the traffic flow on a city-wide level and aid in law enforcement. However, there are challenges in today’s computer vision infrastructure to meet the growing need for compute performance.

The algorithm for each type of object recognition, say, face recognition or vehicle license plate recognition, could be different and thus may require customized and specific compute infrastructure to maximize performance. The common practice is to use a dedicated application-specific hardware for different algorithms which presents a challenge for optimizing compute infrastructure efficiency and system performance. Furthermore, the compute infrastructure must address any surge in compute needs due to heavy traffic in the peak hours. To accommodate peak demand, compute resources are typically over-provisioned and then lie idle in the off-peak times. This is inefficient as well as expensive from both CAPEX and OPEX point of view. For example, in the middle of the night when traffic is usually low. Nothing is moving, so spending expensive compute power to monitor non-moving images is not an efficient use of resources. Same goes for when traffic is holding at a red light.

Observation:The frame doesn’t change for a while, or a large portion of the frame doesn’t change, e.g., time of the day.

Question:Do we really have to waste expensive CNN acceleration resources on those non-moving frames/parts of the frames?

Cascaded heterogeneous accelerated vision cloud (Smart City) Using Software-Defined Acceleration

So, how to address these two challenges? Other than the traditional over-provisioned approach, time-based scaling (see figure below) is commonly used in modern cloud datacenters to provision compute resources based on historical workload pattern. This approach represents an improvement, with more granular resource provisioning but the overall compute resource utilization is still less than optimal. Real time scaling (see below), however, is designed to meet minimum workload demands. Using this approach, different image detection/recognition algorithms can be dynamically loaded, compute resources can be scaled up and provisioned on-demand so that the unused compute resources are off doing other tasks until needed. This keeps the system utilization high, enables much higher levels of efficiency and reduces the Total Cost of Ownership (TCO). For example, at 3 a.m., when there is light traffic and no need for a lot of object recognition due to the reduced number of vehicles and pedestrians, the unused compute resources could be allocated to run alternative workloads, thereby maximizing overall compute resource utilization.

Real-Time scaling for the best TCO

All of this means that software-defined compute acceleration will be at the forefront of Smart City technology and key to bringing all these applications to market quickly. To help facilitate the growing number of applications and drive the trend of Smart City technology forward, NVXL, with its innovative Cascaded Software-Defined Heterogenous Vision Cloud Solution, is partnering with the top cloud service providers to rapidly enable Smart City applications and services.

Want to talk about Vision Cloud tech and your Smart City needs?
Contact me at: and let’s talk

Supporting Resources:

Read about: Smart City Growth

Learn more about: Gartner Smart City Trends

Discover: Top Smart Cities

Find out more about: China’s need for cameras

Stay up to date on: China’s Smart City Development

Uncover: The top 10 Smart City Trends for 2018



Smart City Facts and Figures

Smart City Growth

Untitled Document