Lifetime of a supercomputer

[Idontcare]

This is something I've wondered about for a while - with the extreme level of processing power available in these supercomputers - what exactly do they use them for?

The USA has not physically verified a nuclear device's operational characteristics in 21 yrs.

Everything we think we know about the capabilities of our existing aging nuclear assets is all based on modeling.

Everything we think we know about the functional operating capabilities of anything designed and implemented post 1992 has only been tested in-silico.

Would you be at all comfortable flying in an airplane that had only ever been tested with models on a computer? Never once tested in the lab or in an experimental test flight?

I'd want the models used to be as advanced as the best physicists could assemble, and the computer power put behind those models to spare no expense.

 

[pm]

This is something I've wondered about for a while - with the extreme level of processing power available in these supercomputers - what exactly do they use them for?

The military ones are (or at least were) used to simulate nuclear reactions - in particular, explosions as one might guess, but also deterioration in storage. They create a simulation where they take a three dimensional chunk of the warhead and the air around it, subdivide it out to the computers so that each computer is responsible for figuring out just what happens in that chunk, and then let it simulate what happens.

The NCAR one in Wyoming that I toured did the same thing, except they did it with the weather. They take a chunk of an area and subdivide it out so that various computers are responsible for various locations on the ground and in the air. One of the interesting things about the NCAR was that they allow anyone from an accredited university to submit a request to use the system which is then approved by a committee.

One of the other things about NCAR that I learned is that a lot of the code run on military and weather computers is written in FORTRAN are the product of decades (and decades) of various people coding it and so they really like platforms with excellent FORTRAN compilers and they are not super keen on re-coding anything into other languages, like CUDA for example, although for key portions they will re-code to optimize it. But really didn't like re-coding stuff and so I thought this was interesting that one of the key points of what architectures they prefer is how good the FORTRAN compiler is.

Also one thing that they mentioned was fabric utilization - which essentially is the ethernet network - is how they have one computer talk to another and they said that this is generally the bottleneck in simulations - so not memory, not integer or float throughput, but the ethernet fabric. They quoted some terrible number - and I'm not going to repeat it because I'll get it wrong, but it was pathetic - for how much of the performance "loss" was taken up in the ethernet fabric. They were still working the kinks out and their number was very low and it was their top priority to fix.

They said that the old supercomputers are not destroyed but are broken up and redeployed to other government sites - he said there's some sort of internal Ebay-like thing where other organizations can request and bid on equipment.

One last thing that I have to say is that the tour was amazing because the guides answered any and all questions frankly, honestly and with great technical accuracy - and having worked in the industry with secretetive companies, the military, and even within my own company, I'm not used to people answering all of my questions and not to have any hedging or dodging of any sort... it was pretty cool and very refreshing. If you ever are driving along I-80 in Wyoming and want a break, go check out the NCAR.