If Climavision isn’t on your radar just yet, that’s understandable: the company launched from stealth just six months ago, emerging in June with a formidable $100 million in funding. Its promise: to roll out a combination of numerical weather prediction (NWP), AI, traditional weather observations, satellite data and a new radar network to challenge the major weather prediction players. Normally, this kind of Herculean task might carry with it a large on-prem HPC presence—something seen at many of the organizations that perform large-scale weather forecasting. But for Climavision—fittingly—it’s all in the cloud.

“What we’re trying to do here is replace the need,” explained Jon van Doore, CTO of the Louisville, Kentucky-based company, in an interview with HPCwire. “I mean, these calculations that we’re running for climate modeling used to run on massive Crays.”

“We bring in a bunch of public and proprietary datasets,” van Doore elaborated. “So basically, every weather observation you could possibly imagine gets sucked in, and that’s in the high terabytes. … We crunch it, we spin up this model a few times a day, [it] runs for about an hour, we spin it back down, we then index all that stuff back into our data lake and then we serve it out to APIs and so forth.” Climavision plans to ramp from doing this four times a day—which they say is standard for weather forecasters—to about once an hour, each time generating around a hundred parameters for nearly every point on Earth.

Obviously, this imposes computational costs, and Climavision is employing a bevy of cloud resources to sate those needs, including resources from both Microsoft Azure and AWS. On Azure, Climavision is employing the “latest and greatest” HBv3 instances with AMD Epyc CPUs. “We’re very picky about the sorts of instances we can use and the sorts of processors that go into those instances and how they spin up and down,” van Doore said.

Previously, this sort of cloud-based weather prediction work wouldn’t have been feasible. “It really is the fact that we’ve got high-core count instances, the interconnects now have really good performance—we’re on InfiniBand-backed instances—and the storage now … trying to build out high-speed storage like Lustre or something like that wasn’t even feasible back ten years ago,” explained Brian Dale, director of HPC for Climavision. It was only in the last three of four years, he said, when that changed.

Storage and interconnect, specifically, are key to what Climavision is doing. Once Climavision’s radar network is expanded, van Doore said, the company is “going to be pulling in hundreds of terabytes a day” and pushing out upwards of a petabyte on the other side of things—all of it constantly changing data, unlike major services like Netflix. “That’s a networking challenge that seems to make a lot of these guys sweat on the public cloud teams,” he laughed. “It’s not something that they tend to handle every day.”

These massive data transfers also mean that the cloud instances have to be colocated for the runs to work correctly. Sometimes, a run would spin up and some instances wouldn’t be colocated, destroying the entire run. “That’s why being in lockstep with these providers is so important,” van Doore said. General fragility resulted in the same impacts: if one of 6,000 cores doesn’t spin up correctly, the workload doesn’t run correctly. “So you have to shut the whole thing down, burn it to nothing and then spin it all back up from the ground up,” van Doore said. “And that’s one of the truly frustrating things about the cloud that we’re slowly learning.”

The scale has also been difficult. “There’s a large footprint running in Azure where we have our core instances and so forth set up,” van Doore said. “These things are … powering our normal climate modeling activities day-to-day.” Currently, Climavision is using around 6,000 to 10,000 cores for its bursts—but the company is looking to serve niche customers with separate, bespoke forecasts, and once that ramps up, van Doore says the need will jump to 40,000-60,000 cores “rather easily.”

“Turns out, provisioning these things is expensive and a little tricky!” he said, explaining that Climavision had already used up its allocations in its Azure region and was having to look further afield for cloud resources. Luckily, those bespoke runs allow Climavision to breathe a little easier vis-a-vis colocation concerns. “The reason for the somewhat spread-out, sporadic footprint [of our cloud resources] is because we’re running different workloads in different spots where it makes sense,” van Doore explained. “When we have another workload come out of the woodwork we’re able to say, ‘oh, well, since this is a subset of our work, we can easily section it off and see how it goes.’”

Climavision is working to green up its cloud resources as it expands. “We’re also looking at a really interesting carbon-neutral cloud computing environment,” van Doore said, discussing a provider that powers its instances with natural gas flares. “It wouldn’t be very fair of us to destroy the climate that we’re trying to model!”

Right now, Climavision’s product is in an “intensive build phase,” but they’re looking to have an end-to-end workflow in place around Q2, including an expanded, proprietary radar and satellite network that enables an “occultation-enhanced model.” “Hopefully,” van Doore said, “we’ll get better and faster answers than everybody else.”

Of course, Climavision isn’t the only one gearing up for a cloud-heavy season for weather forecasting: the massive new supercomputer for the UK’s Met Office—billed as the world’s most powerful dedicated weather and climate supercomputer—is being built off-prem by Microsoft and will be Azure-integrated.