How to Leverage a GPU Server for AI and Automation Workloads

What is a GPU server? 

A GPU server taps into the processing power of a dedicated GPU (Graphics Processing Unit) to process large datasets quickly and efficiently. GPUs were designed for parallel processing and were originally created to accelerate the rendering of 3D graphics for immersive gaming and other video capabilities. This parallel computing lends itself well to graphics heavy workloads like image recognition, as well as accelerating the training and inference of neural networks for deep learning algorithms. 

GPU servers are finding applications across a wide variety of industries to help manage the increasingly large amounts of visual data being collected. When you need to power artificial intelligence (AI), machine learning, and automation workloads at the edge, that’s when you need the power of a GPU server. 

Example: AI visual inspection

In manufacturing, visual inspection helps ensure that the products meet specifications and quality standards. Businesses are using AI for visual inspection to catch production anomalies that may be easy for human inspection to miss. 

Mini server

Consider a manufacturing facility looking to implement a quality inspection system right on a production line. The environment might be exposed to temperature variability, dust and vibrations which is why a rugged computer such as the Karbon 700 (K700-X2) would be a good choice to power the computer vision. Check out our customer story about Artemis Vision to see how they implemented the Karbon 700 as a mini-server for one of their vision inspection solutions. 

Artemis Vision creates inspection systems for quality.

Rackmount edge server 

For larger production needs with multiple lines that each require inspection, you could have an industrial PC installed with each production line. Alternatively, all lines could be monitored and controlled by a single high performance rackmount edge server. The high performance computing of the edge server can support multiple production lines with a single piece of hardware. 

GPU Server

To take that a step further, let’s say you had even more complex workloads at each line that required deep learning analysis or other advanced algorithms to be run. You could push your data to the cloud for analysis, but that introduces additional latency. In addition, the cloud can be costly depending on your bandwidth needs. 

To avoid latency and operational costs, many organizations are leveraging cloud repatriation strategies to move their compute resources on-premise. An edge server configured with a GPU has the processing power to meet the requirements for this scenario.

GPU edge server for automation workload

When it comes to automation workloads such as a SCADA system server, you can leverage an edge server with a GPU to process even more tags and connect more HMI clients than you previously could. Not only do these high performance edge servers offer you the ability to run SCADA solutions such as Inductive Automation’s Ignition platform, you also have the ability to enhance these applications further because of the increased GPU compute. 

For example, you could take time series PLC tag data directly from the historian database and feed it into an AI driven predictive maintenance program taking advantage of the onboard GPU. That also has the added benefit of reducing cloud transmission costs and keeping confidential data on-site. 

You also have the option for additional management tools such as creating a highly available Kubernetes cluster for critical operations. 

Installation environment considerations

Edge computing workloads come in all shapes and sizes and installation environments can vary widely. This is why edge computers come with different specifications and in a variety of form factors from fanless or fanless hybrid computers, mini-servers, rackmount edge servers and GPU servers. 

A fanless PC operates reliably in tough environments. That makes it a good choice when deploying in harsh conditions that may include dust, wide temperatures, impact forces and vibration.

Most edge servers require some cooling and dust protection. OnLogic offers high performance edge servers designed for the edge such as the Axial AC101 Edge Server. This shallow depth 1U server offers a number of features for AI and automation, with an operating temperature range of 5° ~ 40°C (ASHRAE A3). However, it still requires a semi-controlled environment to take advantage of that level of compute at the edge.

In some cases you will have the option to deploy hardware in a communications closet or local server rack where it will be protected from the elements. But if you need an edge server directly next to equipment for extreme low latency applications such as high-speed robotics, there are enclosures available that will offer cooling & dust protection. 

As another option, you can look into mini servers that offer many of the same compute features as traditional rackmount servers in a smaller form factor that’s designed to survive in challenging conditions. 

What features should you consider for an edge server or GPU server?

Deployment at the edge introduces some new requirements not as common in traditional server deployments.

CPU performance at the edge

CPU performance is one of the first considerations most people make when looking at a server. Some applications will require Intel® Xeon® scalable CPUs. For other applications, modern Intel Core™ CPUs, which integrate many of the features of previous generation Xeon scalable processors, offer a more affordable, but still extremely powerful, alternative. Axial AC101 offers the latest Intel 13th Gen Hybrid Architecture processors with up to 24 cores and 32 threads.

GPU compute and space limitations

For AI workload acceleration, a PCIe GPU is usually a primary requirement for performance. The AC101 is a short-depth edge server that is capable of supporting a full height, full length GPU. 

1. • This is important as many rack enclosures in industrial environments will not support full depth servers. In addition, you may be limited by the total number of available rack units. A 1U may be your only option once factoring in a UPS, network switch, and other equipment.
   • There are various frameworks that exist to take advantage of GPU compute for AI, such as Nvidia’s CUDA framework which primarily runs on their GPU architecture and Intel’s OpenVino platform which can utilize CPU, iGPU, and Intel based GPUs.

Baseboard Management Controller (BMC)

A BMC allows you to control the server while out of band, which can be key in a distributed architecture at the edge. At scale, effective device management becomes critical. The AC101 does not require users to pay any additional licensing fees to utilize the full capabilities of the BMC unlike many other servers on the market. 

Tamper protection

At the edge, systems can’t always be deployed in areas with access control. To address this, the AC101 comes with a locking security bezel and limited front facing I/O. This helps prevent access to the 4x internal U.3 (NVMe) or SATA SSDs and 2x M.2 SSDs.

Ready to learn how you can leverage a GPU server for your AI and automation workloads? Explore our entire selection of computers available with GPU and reach out to our team today!

Editor’s Note: Inductive Automation has ended their Ignition Onboard program. Ignition licenses must now be purchased directly through Inductive Automation. While the IGN versions of our solutions are no longer available, our computers remain a great fit for use with Ignition software. Explore our recommended hardware here.