From server farm to table: How Nature Fresh uses AI and CPUs to improve crop yields

AI can seem magical, but it certainly takes more than a wave of the wand to leverage AI in the enterprise.?

“There is no magical god for AI,” Keith Bradley, VP of IT for Nature Fresh Farms , told delegates at AI Field Day 4. “We learn more from failures than from successes,” Bradley said of his company’s journey to use AI to help grow 250 acres of crops in greenhouses.

Bradley explained that Nature Fresh Farms has been using technology since the company’s start in 1999. The company aims to take computerized control over the lifecycle of plants to maximize crop yield within the confines of their greenhouses. With 25 years years of data, the company was keen to apply AI to its data, and now uses over 32 different models to help grow bell peppers, tomatoes, strawberries and other crops.

Data comes from IoT devices and cameras that gather data about temperature, soil moisture, humidity, light and more. The data is used by AI models to help control lighting, ventilation, power management and other systems. By closely controlling the growing environment, the company has been able to increase the yield of a plant by threefold.

A few points about the AI the company has in production:

• Models range in size from 4G to 32GB.
• An estimated 60% of the models are “homegrown” and proprietary to Nature Fresh Farms, with the remainder having been developed as an offshoot from other sources.
• Computer vision is used in packing warehouses to ensure that produce of the same age and ripeness gets packed in the same box, minimizing waste and spoilage once it arrives in stores.

Key takeaways

Two key points might go against the popular conception of AI: AI isn’t always in the cloud and AI doesn’t always need to use GPU resources.?

In the case of Nature Fresh Farms, AI inferencing is done at the edge using Intel Xeon CPUs for compute. In this case, the processing is done at the greenhouse or warehouse instead of the cloud because connectivity isn’t a given in the areas they are located, according to Bradley.?

Bradley says there are currently seven Intel-based nodes in use, two of which are dual 4th-generation Xeon. While new compute nodes have GPUs, a separate license would be needed to fully enable them; Bradley says for their current scale, it is not clear that the cost of the license versus what they can get in increased yield would provide a clear benefit.?

What is clear is that the availability of IT systems is essential.? “An hour or two of downtime could lose a whole crop and millions of dollars in damage,” Bradley noted.

The presentation was delivered on day 2 of AI Field Day event, which was sponsored by Intel. Collectively, Intel customers and partners discussed the types of AI inferencing workloads were appropriate for CPUs.