Embedded Editor Report: AI at the Edge Shines at NVIDIA GTC

March 20, 2024

Story

The story of the second day at NVIDIA GTC in San Jose is Embedded AI for Edge, at least that was the story I was hearing again and again from the folks that I talked to on the show floor.

Of course, there were amazing innovations on display in robotics (that’s the edge, though, right?), automotive, especially Software Defined Vehicles (still Edge, I’d say), Physical servers (not the edge this time…or is it?), and many other areas. But, joking aside, it seemed like everyone I talked to either wanted to get AI working for them managing data and providing intelligence in their Edge devices and networks, or they wanted to provide the embedded computing that was purpose-built to make that happen.

As you might imagine, the floor was busy as these two disparate groups sought and found one another. An AI industry meet-cute of sorts.

Let’s dig a little into just a few of the conversations I had with some of the companies creating these Embedded Edge AI solutions and you will see what I mean.

Advantech

Obviously a huge player in its own right, Advantech was showing off some of its products and solutions developed specifically in partnership with NVIDIA and based on Jetson. There were many things to see in the booth, but of particular interest was the video-to-text solution in which an AI analyzed a video feed in real-time and then described what it was seeing. As the scene changed and different people or items come into focus and foreground, the description adapted in real-time and described the new scene. The latency was very low, and the AI was usually very accurate, and when it made a rare mistake, it self-corrected quite rapidly.

Arbor

Arbor is a company offering what they call custom solutions for edge processing at lower latencies. The company’s AEC-6100, FPC-5211-M4, and FPC-9108-G3 models on display in their booth showcased a range of offerings across a range of customer needs for I/O, ruggedization, and power, but all boasted low-latency processing of data, suited to AI-enabled applications, especially for video monitoring and sensing (think computer/machine vision), and in vehicles (especially the powerful FPC-5211-M4).

Delta

As the only company on the floor doing power supply in partnership with NVIDIA, Delta is making power conversion smarter and more powerful. The company is showing off its new ORV3 racks, the 33 kW power shelf, and the company’s board-level DC/DC Converters VR Series, that feature vertical power delivery architecture to increase TVR efficiency of output voltage for GPUs to over 94 percent.

Innodisk

Innodisk is using AI inference with NVIDIA algorithms to improve industrial inspection systems with AI. The engine compares products on the line to a “golden image” of the product and rejects any found to be faulty against the parameters, and the AI was working so fast, it was actually difficult to keep up visually. The camera and processing didn’t seem to have any difficulty, however. They have reportedly improved accuracy from industry averages of around 70 percent to upwards of 95 percent correct identification.  

neousys

This interesting company is building solutions with SDKs, intended for outdoor or industrial deployments, so they are ruggedized, fan less and temperature resilient, as you might imagine. Most impressive was the waterproofing demo they had on display with a working computer processing video in real time, while a hose blasted it with water. No faults. No lag. No problem.

This is just a sample of the many great AI and Embedded solutions, products, and applications I saw this week, and I’ll be bringing you more over the next few days as I start to unpack some of the implications of everything NVIDIA showed us.

Keep your eye on this space for more!