Designing and scaling industrial computers is becoming tougher as the types of sensors and the number of sensor channels needed to satisfy the exploding appetite for manufacturing data continue to increase.
Furthermore, as industrialists and medical organizations look to take advantage of automation, systems are increasingly infused with AI, ML, data analytics software and intelligent displays. This drives the need for greater levels of diversified computing. New approaches enabled by adaptive compute platforms conceived for sensor-rich control applications can accelerate development, simplify hardware-software integration, and sustain performance while allowing tight control over power consumption.
Embedded PC computing trends
The ongoing digitization of edge applications comprises several elements including sensorization, the infusion of AI and machine learning across edge and cloud computing, human-machine interfacing, multimedia experience, networking and integration of Operational Technology (OT) and Information Technology (IT) domains, which often need different compute elements to perform optimally.
Take an example of a medical imaging system. It typically includes probes that need to be interfaced and processed using various algorithms, which require a large amount of computing given the complexity of the workloads. The data created by these operations is only useful to medical users such as radiologists and cardiologists once it has been cleaned, organized and processed. Data analytics engines and AI inferencing can also generate insights to accelerate the process of analyzing the results. All this information must be rendered and visualized on display monitors to aid the medical analysts and cascaded to a medical database via the organization’s network.
This is just one example of a process where extensive sensorization enables a multitude of changes that can enhance efficiency and productivity for embedded applications. These wide-ranging sensors need to be interfaced and processed promptly, usually in the range of milliseconds to achieve maximum responsiveness. Massive sensor deployments also feed Big Data algorithms that extract intelligence about processes and generate insights that drive improvement and next-generation product development.
Large numbers of edge instances that have sensors and adaptive compute platforms also have a PC either incorporated or tethered to their system. Bringing x86 computing, AI, control, sensor interfacing and processing, visualization and networking closer together delivers important advantages, including size reduction that eases deployment and installation. In addition, savings in power consumption simplify power supply design and can permit battery-powered applications such as AMRs used for moving components, materials, and finished products within the factory to operate for longer on a single charge. Whereas fitting a larger battery to achieve the runtime lost due to excessive power consumption in the processing engine adds to the cost and weight of the system, a more integrated solution permits a lower cost of ownership.
However, integration demands intensive hardware and software engineering effort, which continues as more and more sensor channels are added in pursuit of greater productivity, safety, and more efficient business planning.
Flexible integration
A common approach is to take advantage of the rich ecosystem around the x86 processor architecture, commonly used in industrial and medical computing, alongside an adaptive compute platform that can execute real-time machine control, sensor interfacing and networking. This combination can be applied to use cases such as machine vision, industrial networking, robot controllers, medical imaging, smart city, security and retail analytics.
Conventionally, an industrial PC acts as a gatekeeper, dealing with the influx of sensor data and arbitrating whether processing will be handled on the x86 core or, if available, an FPGA-based accelerator card accessed across the PCIe® interface. Latency is a major problem encountered with this approach. The time to ingest, process, and transfer the sensor data into the accelerator adds delays that can make real-time system response impossible.
The proposition of integrating sensor interfaces, AI processors, and network processing into the FPGA-based adaptive computing platform holds immense promise. Consolidating these functionalities onto a single motherboard enhances computational efficiency and reduces latency, eliminating the need for data to traverse disparate components. This integrated approach offers the potential for faster response and greater accuracy, as well as lower power consumption.
Supportive ecosystem
Adaptive compute platforms that can handle real-time sensor processing, control, networking and AI inferencing can help minimize latency, power, and overall solution size. The result creates an efficient and powerful platform for embedded processing.
Extending this principle, embodied in devices such as AMD Versal adaptive heterogeneous processors, can help streamline building embedded compute platforms to support the trend towards sensorization while tackling existing diversified workloads. With the addition of x86 processor IP, leveraging specialist adaptive compute solutions, as well as large numbers of I/O suited to sensor interfacing, the next level of integration, power efficiency, and system response is within reach. Large numbers of I/Os make it possible to connect different types of sensors and route the signals directly to be processed. This can apply to many different types of sensors such as GMSL (Gigabit Multimedia Serial Link) cameras, 10/25GE, LiDAR, medical probes like endoscopes and ultrasound. Moreover, additional sensor channels can be configured relatively easily when needed, thereby assisting scalability.
This approach combines the advantages of scalable sensor interfacing and heterogeneous acceleration with the advantages of the large ecosystem supporting industrial processing on x86 platforms to simplify sensing, AI, and control and networking software. Engineers can thus build the optimized embedded computer that exactly meets their needs. They can tailor the number of sensor I/Os, bring each channel individually into the most suitable acceleration engine whether this is a CPU, real-time core, DSP, or AI engine, programmable logic and fine-tune the implementation for optimum power consumption as well as performance. The flexibility to connect the signals on any input channel to the most suitable processing engine on the chip also helps engineers handle mixed sensor criticality, according to importance and dependence on real-time determinism.
The extensive ecosystem supporting x86 embedded computing provides rich resources to support the development of applications such as machine vision, medical image scanning, robot control, and others.
Written by KV Thanjavur Bhaaskar, Technical Marketing and Strategy Manager, Industrial, Vision, Healthcare & Lifesciences, AMD
