DigiFarm´s core operation consists of the development of technology and model training of deep neural networks that can automatically detect field boundaries, sown area, and other objects within land changes such as trees, bogs, shade (from trees), telephone poles, irrigation divots. The models are trained in several international regions including Europe, South America, APAC (Thailand, Vietnam, and India) as well as the USA and Canada.
Artificial intelligent agriculture
This requires a large amount of image data as well as annotation data for training models as they process from over 300 million hectares of satellite data with 1 m resolution every month from 3 different dates of data, as well as to detect 30 million hectares with land changes.
According to CEO and Founder, Nils Helset, this requires a stable, secure, and high-performance HPC infrastructure with GPUs. DigiFarm has been part of several cloud-based startup programs including Google Cloud for Startups (A100 alpha program), AWS Activate, AWS Open Earth Initiative, Microsoft for Startups, NVIDIA Inception, and Oracle for Startups, but each of these platforms and programs has several challenges:
- GPU costs - the average price of GPUs (NVIDIA V100 and P100) is approx. $ 2.75 per hour - we currently use 35 GPUs constantly for our development where this time price becomes too expensive for our operations, especially with regards to there being restrictions on the type of GPUs you have access to free cloud credits and startup programs, in Azure (Microsoft for Startup) you only have access to K80 while in AWS, Google and Oracle then A100 is not available, only V100 and P100.
- Workshop and technical support - because we have a complex computer process infrastructure with many variables and different elements, it is incredibly useful for us to have a continuous dialogue with the National Competence Centre for HPC that helps us with the optimization of our processes.