This kind of visualization ability generated through open source GEN AI technology is groundbreaking in its own right. But the setup behind that tool is just as innovative.
Consider your calculation power. Depending on the band and mosaic process, global satellite images can be clocked at 3298 x 9896 pixels (and more), and 15 years of data measured every 30 minutes produces 263,000 images. This is 17 TB of data per training session for the GAIA model. The team also manipulates live weather data, leveraging the same operating satellites that weather forecasters use in the news at night. These basic model approaches require a Many This is the common reason why visual-based GEN AI tools traditionally have fewer exploration spaces among GPUs.
“Until now, pure computation, algorithms and know-how that needed to be able to translate pixels into answers has been extremely rare,” says Potte.
Tackling calculation problems
BCG X made two conscious decisions when monitoring efforts so that the project could be delivered online in just one year, rather than the 18-24 months typical for other projects.
The first was to create an environment that could be deployed in the cloud, rather than connecting to a dedicated supercomputer. According to Tom Berg, BCG X lead engineer at the project, “There was something really difficult here. Rolling up your sleeves and building your own basic model is almost expensive to build your own basic model, looking at the immeasurable resources that hyperscalers use.
To that end, the GAIA team turned to a national network of university computing resources distributed to the US. This constellation of this ready-made GPU (a range of cutting edge GPUs to 10-year-old GPUs) is exactly what the BCG X development team had in mind.
“That profile provided many parameters to work with, rather than matching the supercomputer,” says Berg. “It was a very adaptable system, and at one point we were using 15% of the entire NRP cloud.”
Still, such a setup provided some interesting challenges. If a dedicated supercomputer has all the processing power in one building with one uniform power configuration, the Berg and Potere teams will instead connect the GPU on the other side of the globe. There were also acute issues such as power outages, and universities unexpectedly cycling data centers. Importantly, Gaia shared computing space with hundreds of other research applications running simultaneously. “You're basically on busy public roads, not on a dedicated racecourse,” says Berg.
The team's second operational decision was to focus first on precipitation and top temperature data, as opposed to modeling all aspects of every layer of the atmosphere. As the selected data closely corresponds to a variety of weather phenomena, researchers provided the flexibility needed to prove the basic model and carry out the experiment with a level of management-level initial effort.
