How heterogeneous computing optimizes deep learning workloads

As your manufacturing operation begins to explore the value that deep learning can offer, you’ll notice that it gives you the possibility to use many different types of data. Your business won’t continue to just use data collected by barcode readers or inputted on a keypad. Deep learning applications can extract value from a variety of data types including images, video, text, speech, and sensors. Because data comes in many, heterogeneous forms, deep learning applications need heterogeneous computing which uses different types of computing cores to process data.

What is Heterogeneous Computing?

Heterogeneous computing uses two or more types of computing cores, such as:

  • Central processing unit (CPU)
  • Graphics processing unit (GPU)
  • Field programmable gate array (FPGA)
  • Application-specific integrated circuit (ASIC)

Using multiple cores provides systems with capabilities that single-core processes can’t match. For example, a system that employs both FPGAs and GPUs not only quickly render images, but they can also rapidly perform multiple calculations at the same time, while FPGAs enable in-field programming. Deep learning can also combine these cores with various types of ASICs — vision processing units (VPUs), used for computer vision, and tensor processing units (TPUs) to provide specialized capabilities with low power consumption.

Heterogeneous computing is not a new concept; different types of computing cores are already working together in a variety of applications. Deep learning platforms, however, requires integrating different types of computing cores into one system to deliver optimal performance per watt. Using the wrong cores or designing a system with computer core “overkill” would prevent deep learning applications from being optimal, in terms of size, weight, and power consumption.

What Heterogeneous Computing Can Do

It’s possible that, to some extent, your operation is leveraging systems built with heterogeneous architecture now, and it’s probable that heterogeneous computing will play an essential role in reaching your Industry 4.0 objectives. Using multiple types of computing cores in a single platform enables a wide range of applications such as:

  • Machine vision for more efficient quality control inspections, measurement, and enhanced process control
  • Optical Character Recognition (OCR) for a variety of use cases including extracting data from legacy equipment
  • Autonomous mobile robots, which can increase the speed of routine operations and perform tasks that pose a risk to employee safety
  • Internet of Things (IoT) environments that stream and analyze data in real time between devices and systems, resulting in greater efficiency and better decision making
  • Deep learning systems that enable algorithms, first to learn and then to infer directly from images, speech, or video, without requiring you to tell it how to do the job.
  • Parallel computing that can accelerate image processing and, therefore, deep learning inferencing.

The Advantages to Your Manufacturing Operation

One of the most significant advantages of heterogeneous computing is that, by integrating different types of computing cores into a system, you can optimize performance per watt as well as per dollar, size, weight and power (SWaP), as these parameters must be taken into account when deploying edge AI solutions.

Heterogeneous architecture also allows you to scale your computing environment more easily. For example, with ADLINK’s heterogeneous computing platform, with include the Vortex Data Distribution Service and microservices architecture for external workload consolidation, you have the ability to expand your AI system to include new technologies as they emerge and enable new opportunities for innovation and competitiveness.

Don’t Limit Your Potential with a One-Track Mindset

Separate systems each built with a single type of computer processing core will limit your operation from building the AI-enabled, Industry 4.0 environment you envision. Your potential to have visibility into your operations, prevent downtime, maximize productivity with remote control and automation, and enhance worker safety will not be competitive with systems leveraging heterogeneous computing.

Want to learn more? 
Join us at NVIDIA’s virtual #GTC20 conference to hear from our Edge AI experts with registration promo code NVESEYMOUR!

And download our latest heterogeneous computing white paper

Author: Zane Tsai
Author: Zane Tsai

Director of Platform Product Center, Embedded Platforms & Modules, ADLINK Technology