Defining the Right Amount of Compute Power for AI at the Edge

By Flay Mohle

Director of Marketing

WinSystems, Inc.

April 19, 2022

Story

Defining the Right Amount of Compute Power for AI at the Edge

We’ve spent lots of time discussing the Edge of the IoT, particularly what processes should be handled at the Edge versus those that are better performed in the Cloud. And you’ve probably noticed the focus moving increasingly more toward the Edge, as compute capability grows.

But that discussion soon leads to the next point: just how much compute power do I need to perform my operations at the Edge? This is a valid question, for several reasons. For example, more processing power is simply more expensive, and you don’t want to over-design and then pay more for the system than is necessary.

Next, more compute capability generally requires more power, and that brings with it its own set of issues. That discussion starts with more heat that needs to be dissipated, more fans that add noise, additional moving parts, and a larger footprint for each embedded computer that’s deployed.

Finding the Sweet Spot

The amount of compute power needed comes back to what your specific application is. In every case, you want just enough, but not too much, for the reasons explained above. The only instance for considering more compute power than necessary would be to future-proof your design, if you’re confident that you’re going to need more compute power in the near future.

 

The WINSYSTEMS’ PX1-C441 SBC boasts ample compute power for industrial IoT applications and embedded systems in the industrial control, transportation, Mil/COTS, and energy markets.

For example, you could start with a WINSYSTEMS PX1-C441 single board computer (SBC) and have most of your processing needs covered. The SBC is designed according to the PC/104 form factor and includes PCIe/104 OneBank expansion. PCIe/104 OneBank takes advantage of a smaller, lower-cost bus connector that’s compatible with the full size PCIe/104 connector which is designed into millions of embedded computers today. It lets designers stack boards using a complimentary format that frees up board space for additional components.

Developers will also see gains from the PX1-C441’s Apollo Lake-I dual- or quad-core SoC processors. This latest-generation CPU family is suited for Edge-based IoT applications, including those that require graphics processing. This level of compute power, combined with small size and a rugged design, makes the SBC suitable for industrial IoT applications and embedded systems in the industrial control, transportation, Mil/COTS, and energy markets.

Other features of the PX1-C441 include up to 8 Gbytes of soldered down LPDDR4 system memory and a non-removable eMMC device for solid-state storage of an operating system (OS) and applications. In addition, the board supports M.2 and SATA devices.

Many Variables to Consider

When deciding upon which CPU to employ in your IoT network, aside from the straight compute power which is usually defined in MFLOPS, a host of features must be considered, and they vary by processor vendor and family. The features you should be looking at include processor architecture, memory footprint, power consumption, communication protocols and ports supported, security, existing software code base, hardware and software ecosystem, and of course, cost.

Oftentimes, feature sets are bundled, and you can’t get exactly what you need, but it’s usually close enough. While embedded computer boards can often be customized for a specific application that’s not the case with microprocessors themselves.

Architecture typically refers to the instruction set that runs on the processor, whether it’s X86 (Intel or AMD), Arm (a host of vendors), NVIDIA, or something else. And the memory footprint can easily scale up or down, as needed. Power consumption is a hot button (no pun intended) and it can have negative effects further along in your design, as we discussed earlier.

And don’t underestimate the ecosystem or partner network of your vendor. A great example is the WINSYSTEM’s partner network, an ecosystem that can deliver the solutions you need to fit your application, both in terms of hardware and software. WINSYSTEMS vets the ecosystem partners to ensure that all their offerings work as advertised on their boards with little to no tweaking.

As you can see, there’s no simple answer to the question of how much processing capability is enough. But if you align yourself with the right partner, like WINSYSTEMS, they can help guide you to make the most educated decision.

As marketing manager at WINSYSTEMS, Mohle leads all branding, strategic marketing, creative advertising and corporate marketing communications activities. He also manages outside agency partners for website maintenance, SEO/SEM, content generation, email marketing, press relations and industry trade show activities.

More from Flay