The world's fastest computers

Typically thee machines are actually running lots of large problems at once.

So, if I give an example, here in the UK, academics have access to a few national grid based services (not the same as these machines, which although distributed, are done so in such a way that inter-connects dont effect their overall operating speed as much), these national services offer in the region of around 10,000 cores, we then book time on a chunk of those cores, normally a few hundred.

You then go away, work out how to parallelise your problem over a few hundred cores (its suprisingly difficult for a lot of simulations with only a few being what we call embarassingly parallel), and when your time comes, you submit your problem as a batch to run over a specific time-period, people normally get allocated a few days to a week.

So while on paper these machines have mammoth peak processing power, much of the time the jobs being run on them are using only a fraction.

Earthquake modelling can be stupidly intensive (not difficult per se, just time consuming).
 
It would be fairly powerful, yet consume more power than the sun.

Mmm, top of the range processors are typically utter crap when it comes to efficiency/watt and multiplied by tens of thousands you're looking at having a whole dam substation built to power it and a black hole of an electricity bill.

Atoms should be pretty good at this tbh, wonder when someone will make an atom based supercomputer :)
 
Intel account for about 81% (402) of the top 500 supercomputers in the world.

AMD account for about 8.5% (42 supercomputers) but crunch about 30% of the total amount Intel do.

AMD happen to be able to sell their chips for a very low price/core. It's hard to argue with that on supercomputer scales.
 
Intel account for about 81% (402) of the top 500 supercomputers in the world.

AMD account for about 8.5% (42 supercomputers) but crunch about 30% of the total amount Intel do.

AMD happen to be able to sell their chips for a very low price/core. It's hard to argue with that on supercomputer scales.

And 90% run on Linux.
 
Its less about cost, but scalability, something that for a few years, AMD has done incredibly well on, Intel might have QPI and on die mem controllers and trip channel memory NOW, but you don't just build a supercomputer over night, planning time and design all takes a while. In another 2-3 years some of the more complex and huge new supercomputers will be based on i7's/i9's, and in another 3 years they'll be bulldozers.

Though theres one thing, AMD might still dominate, depends exactly what platform those opterons are on, but AMD will be offering more compatibility for dropping in a quad/6 core than Intel do, meaning several of them could be upgraded to be far more powerful with little more than new chips.

Most of them are still lots and lots of dual core chips, because as said they are general based on not completely up to date equipment. Its almost surprising the China one uses 4870's as they are still fairly new, infact brand new in terms of supercomputer stuff.

The newest top 5-10 supercomputers every year will always do a disproportionate amount of number crunching compared to the next 10 and the rest after that as there aren't that many supercomputers.

Very interesting to see though, I don't know what would be more surprising, how much they cost to put together, or how much cash they burn through in power. Considering they are often cooled by a ridiculous amount of AC units aswell.
 
Back
Top Bottom