Italy's climate super computer, Cassandra, to combine HPC with AI Boffins in Italy are about to get their hands on a supercomputer that will more than double the resources available to study the effects of climate change. Dubbed Cassandra, the big iron is based on Lenovo's liquid cooled Neptune system architecture and is slated for deployment at the Euro-Mediterranean Center on Climate …
I would have thought that a climate modeling supercomputer would want to be extremely efficient energy usage wise, though I'm not surprised to see it not mentioned.
I'm sure there's no good way to spin your local area losing power due to this thing and its ilk using so much, only to have to explain that the coal/oil power plants could not keep up with demand and admit the renewable energy is no where near close for this application.
All while spinning up gas fired back up generators to run a computer trying to find a way to do less damage to the climate...
Just that its conspicuous by its omission. Seems to be a metric in many other supercomputer posts. Why not this one?
And its a legitimate concern, we know having them in local built up areas is not a great idea. So why Italy, while they do have the alps, most of it is know for being sunny and warm. Cooling is a large part of the power requirements.
There are location in Europe far better equied to do what this needs but in a more environmentally freidnly way. Natural water cooling? Cold air intake from naturally cold areas. Areas with surplus renewable energy frequently.
It would seem odd to me that with its intended purposes every effort was not made to make it as efficient, cool running and supplied by as much renewable energy as possible.
And at the end of it, we will just confirm want we already know, then offer ideas on how to make a future possible, and have them ignored by the governments and corporations.
"There are location in Europe far better equied to do what this needs but in a more environmentally freidnly way. "
Yes, but not that much. Sure, Lecce is in southern Italy and quite warm, but there are other advantages to working locally - it's also an energy cost to transfer Terabytes of data somewhere cooler to do the calculations and transfer the results back, not to mention budgetary costs. In the end, the energy expenditure of the machine is likely to be many orders of magnitude less than the gains made for any sensible climate policy changes that derive from its results
..it's also an energy cost to transfer Terabytes of data somewhere cooler to do the calculations and transfer the results back, not to mention budgetary costs.
Possibly not that much, but depends on the modelling. Basic modelling is a wicked problem, but relatively easy. Break the world up into a bunch of grid squares, add maybe half a dozen parameters, add in the calculations and call it good. Then you 'just' have to decide how often to run the numbers, pass the data and hit 'Go'. So maybe a few million calculations and parameter passing per simulated hour and you'd like the results for say, 100yrs into the future and you really want the results back faster than real-time. So with bigger iron, you can do more runs per work day, tweak more parameters and generate more results.
Inputs for that don't need to be Terabytes, and neither are the outputs.
Where it gets more complicated is doing actual weather forecasting, because that's far more complicated. There, you're trying to accurately predict weather, not climate because climate is just averaged weather. Plus (from memory when I lived in Reading and drank with ECMWF folks) you need to produce 3 models a day because people rely on those forecasts. That requires ingesting far more data, especially as the 'climate hype' has had the bonus of producing more weather observation systems and data sources. Model output isn't that large either.
Then it gets a little more complicated if you're doing reanalysys, ie comparing models to reality.. But that still doesn't need that much data given you may only be comparing hourly temps, precipitation, relative humidity etc. But there you're constrained by both availability of data, and number of grid squares in your model vs observations.
I think the architecture questions get more interesting, but are waaay outside my area of expertise. I have however had some fun chats about this with people who do this stuff. So I've always been curious about CPU vs GPU given most of the world is analogue and vector performance just seems better. That also touched on aspects like FPC64 vs 32, or even smaller. After all, if your data are often all capable of being represented in 8bits, do you really need 64bit performance? So for example temperature data can only be measured to maybe 2 decimal places, and often with high error margins even then.
Which was all rather fascinating. So stuff like could CUDA cores on a GPU act as 'cells' for models, and pass results on to other cores more quickly than a CPU that's perhaps better optimised to calculate in serial rather than parallel.. And generally the answer was 'yes', but it all gets very complicated very quickly, and alcohol was involved.
> So why Italy, while they do have the alps, most of it is know for being sunny and warm. Cooling is a large part of the power requirements.
So, no datacenter under the 60th parallel? Sorry guys, abacus is still allowed though.
I hope you were located in Bergen (60°) when you typed this post. Oslo is off limit (59°).
I would bet money that the choice was deliberate.
You would hope so but Cockup out of Incompetence is aways odds on favourite over Conspiracy in the Italian Stakes. Kassandra and horses not a fortunate combination.
Sibyl might have been better, but given their notoriously ambiguous prophesies, perhaps not. Not that Mrs Fawlty suffered from that disability. ;)
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2025
