Met Office ‘supercomputing as a service’ one year old
Artificial intelligence is not key to the weather picture, as the forecasting and climate prediction agency lauds the benefits of moving from on-site supercomputers to cloud computing for scientific modelling
The Met Office has celebrated one year of “supercomputing as a service” from Microsoft.
The occasion presented an opportunity for the weather forecasting and climate prediction organisation to offer its view on why artificial intelligence (AI) is peripheral to its scientific modelling core activities and why cloud is suited to the delivery of supercomputing services.
The Met Office’s Microsoft cloud supercomputing capability was launched a year ago, and provides around 1.8 million processing cores, with peak performance of around 60 quadrillion calculations per second.
Reported availability for the system in the past year has been 100% for critical workloads, with 99.77% for the supercomputing element.
Microsoft delivers supercomputing-like performance remotely via the cloud. But what makes the cloud suited to delivering such services, when latency might be an issue compared with on-site hardware?
According to the Met Office and Microsoft, cloud latency is comparable to what it was between the Met Office headquarters and where its previous supercomputers were located.
In addition, it pointed out that latency is only one factor taken into account when deciding whether to go to the cloud. Others include cost, reliability, flexibility and the ability to take advantage of innovation, which might have been limited with a multi-tonne, multimillion-pound supercomputer lodged in a Met Office facility.
The Met Office’s work is still largely physics-driven numerical weather prediction and modelling. AI is therefore not the lens through which things are viewed, said Met Office CIO Charles Ewen, but one tool among many.
“People are struggling because they’re looking through the narrow lens of AI at the moment,” said Ewen. “And before that, it was cloud. AI is certainly happening faster and quicker. And not to diminish AI and its importance, but we have to think more of AI as a catalyst and an accelerant for broader innovation.”
So, what makes it a supercomputer if it is merely computing delivered via the cloud? Ewens defined supercomputing as typically used to do things like the kind of scientific simulation that is intrinsic to the production of weather forecasts and climate predictions.
“The services we’re consuming today are really enterprise-scale cloud computing for scientific supercomputing. It is certainly in the top five of CPU [central processing unit] clusters in the world. In my assessment, at least, the only truly cloud-integrated scientific supercomputer.”
While AI is currently peripheral to scientific modelling, Microsoft Copilot and generative AI are in use in the Met Office organisation more widely, and are being looked at for scientific work, said Ewens.
“AI methods are coming along, and one of the benefits of the work we’re doing is we are already well into planning how AI methods, data-driven methods, sit alongside physics-driven methods or the more traditional numerical weather prediction methods in the best blend to deliver the very best of both.”
On data sovereignty, it is understood that all Met Office operational datacentres are in the UK, and that Microsoft can supply sovereign cloud capability that ranges from fully connected to the cloud to fully disconnected.
The Microsoft-powered cloud capability is the Met Office’s 14th iteration of its supercomputing capability. Numerical weather prediction began in the early 1950s, when the first experimental forecasts used the EDSAC computer at Cambridge.
By 1959, the Met Office took delivery of its Ferranti Mercury computer at its Dunstable site. The Met Office produced its first operational computer forecast in 1965, following the arrival of an English Electric KDF9 computer at Bracknell.
In the past decade or so, the Met Office went from buying a 140-tonne Cray XC40 system in 2014 – its fourth supercomputer at the time – to opting for cloud-delivered supercomputing in 2024, with the current Microsoft-operated, fully managed “supercomputing-as-a-service” model.
