The UK's obsession with the weather has just been stepped up a gear. The Met Office has announced that it will begin using a new supercomputer enabling highly detailed forecasts. The Cray XC40TM will cost £97 million (US$156 million) and is aimed at making the UK a world leader in weather and climate science.
The Met Office is the UK's public weather and climate research agency. The data it produces is made available for use by other organizations and has previously been employed for showing temperature records in Google Earth and and by Micasa Lab's indoor cloud-making machine. This most recent investment will help to improve the volume and accuracy of the data produced.
The Cray XC40TM is one of the fastest supercomputers in the world. It has 480,000 cores, 2 million gigabytes of memory and can store up to 17 million gigabytes of data. At its peak, it is able to make 16,000 trillion calculations per second.
The system is a significant upgrade from the current IBM Power 775, which has a memory of 80,000 gigabytes, can store up to 2 million gigabytes and can make 1,200 trillion calculations per second.
The computer will enable hourly forecasting with a much higher level of weather detail for precise geographic areas and the ability to create much higher resolution weather modeling for localized areas. The Met Office says that that it will be possible to create models for areas down to a resolution of 300 m (984 ft), allowing it to effectively predict differences in weather conditions between city blocks or even individual streets. It will also be possible to better predict weather over longer periods, allowing for better preparation against developing climate trends such as flooding, droughts and heatwaves.
According to the Met Office, this could lead to £2 billion of socio-economic benefits to the UK, allowing airports to operate at a greater level of efficiency, properties to be better protected from flooding and winter infrastructure to be better planned. Planning for renewable energy infrastructure and climate change research will also benefit.
The Cray XC40TM will be located partly at the Met Office HQ in Exeter and partly at a purpose-designed building at the nearby Exeter Science Park (pending planning permission). It will become semi-operational in September 2015 and will reach full capacity in 2017.
Source: Met Office
On a more serious note, whilst making the UK a world leader in climate science is a laudable goal, unfortunately there is a sizeable portion of the public that still will not believe that the climate is even changing, let alone just how serious a danger it is becoming ("drought, what drought?"). If this new computer could somehow get them to wake up and smell the coffee, it would be worth every penny, even at ten times the price.
@ Justin -- absoutely not! A machine like this is purchased for the interconnect. It can send a message from any core to any core in on the order of 1 microsecond. Think about it 480,000 cores each within 1 microsecond of each other. Systems like this ARE a rack of servers but with the addition of the interconnect it can do so much more. An equivalent sized server farm would do weather prediction 100-1000X slower with the only difference being the interconnect. The details of pricing are not divulged (and I don't know them personally for this procurement) but the interconnect costs ~ 30% of system cost for a 50% increase in cost for a 100-1000X improvement - wise investment and why a server farm simply will not do. But beyond that, each core has access to ~ 4 GB of memory - most server farm or VM images have much, much less memory footprint - again large memory/core is needed for weather but not other std. server farm apps like web or app servers. Not mentioned is the I/O rate, but knowing Cray, I can guess it is considerable somewhere between 1 terabyte/sec to 500 terabyte/sec I/O rate to disk (Iops will be impressive but not as impressive as the BW numbers)
This is not your dad's server farm!
And in terms of cost --- apx. £200/core. Competitive pricing with server farms in terms of server farm language. In terms of service to user, the cost per flop is probably smaller than any other option, except perhaps for some other supercomputing centers.
Left as an exercise for the students - estimate the power needs of this system and the annual power cost for this system.