Summit supercomputer nears completion, will be fastest in the world

William Gayde

Posts: 382   +5
Staff

China currently leads the world in supercomputing, housing both the world's fastest and second fastest machines according to TOP500 rankings. Not wanting to be outdone, the US has been constructing what it believes will be the world's fastest supercomputer upon completion.

Summit, as it will be named, is scheduled to have between 150 and 300 petaflops of computing performance and will be unveiled at Oak Ridge National Laboratory in 2018. The current record holder, Sunway TaihuLight has a performance of 93 petaflops. Summit has been designed as a replacement for the previous top US supercomputer, Titan, which is capable of 17.59 petaflops.

To put those numbers in perspective, modern high-end consumer CPUs can only perform about 50-100 gigaflops and high-end GPUs can perform a few teraflops. This makes Summit thousands to millions of times more powerful than a home computer.

Summit will consist of about 4,600 nodes, each with a performance exceeding 40 TFLOPs. Individual nodes will have 512GB of DDR4 memory, an additional 1,600GB of non-volatile memory and extra HBM although the specific amount isn't specified. In total, the system will contain more than 10PB of RAM. In terms of compute, each node will have two IBM Power9 processors and six Nvidia Volta GPUs. The datasets it will need to digest are equally as massive so Summit will pack 250PB of file storage with a performance of 2.5TB/s.

The full system will consume roughly 15MW of power at peak load. This is the same as Sunway TaihuLight but will offer 2-3 times its performance.

Looking to the future, Oak Ridge also plans to develop an exascale computer by 2021 (that's one billion billion operations per second).

Permalink to story.

 
"Not wanting to be outdone, the US has been constructing what it believes will be the world's fastest supercomputer upon completion."

Is this a US vs China thing, or is because these Summit builders, as a separate entity actually need the power, and will benefit from the power.

Is this the case where some guy in the neighborhood bought a computer more powerful than yours, and even though you don't need the power you're going to buy and even more powerful one just to be number one?
 
"Not wanting to be outdone, the US has been constructing what it believes will be the world's fastest supercomputer upon completion."

Is this a US vs China thing, or is because these Summit builders, as a separate entity actually need the power, and will benefit from the power.

Is this the case where some guy in the neighborhood bought a computer more powerful than yours, and even though you don't need the power you're going to buy and even more powerful one just to be number one?
Oak Ridge actually needs a big upgrade, titan was just a upgrade over jaguar, swapping out k10 cpus for bulldozer based and adding a new generation of tesla cards, jaguar was the best they could with 15MW in 2008-2009, Titan was the best at 15MW for 2011-2012 and Summit represents the massive gains in performance per watt over the last 5 years. As far as tit for tat the Chinese have come a long way in numbers the USA has 168 of the top 500 China now has 160. The USA has dropped around 30 major systems that I have heard of since Universites can rent time on larger more powerful systems they don't feel the need to have those onsite as much, amazing what a 10Gb internet connection can do :D. The US also current holds 5 of the top 10 fastest super computers and we are still like 12-14 of the top 25, so a lot of the most powerful are here in the USA. China upgraded early and got the head start on everyone, Europe, USA and Japan will be investing a lot into mainframes in the next 3-4 years and the list will show the change up.
 
My question is how do they go about saying it could have x (150pflops) amount of power or it could have 2x (300pflops) the power, were not sure. That's double the supercomputer, that is massive gains. I don't see how they can't get a little closer with the guess than that unless they are just trying to puff their chest before the real numbers come in. I'm just going to play it safe and say 225 lol.
 
When's someone going to build a supercomputer out of AMD's latest? (Just curious to see how it will turn out)
 
When's someone going to build a supercomputer out of AMD's latest? (Just curious to see how it will turn out)

The last ones were built out of Bulldozer-based silicon. So much has changed in the last 5 years though that the silicon inside of super computers is nothing like the silicon in consumer hardware anymore.
 
A super computer isn't really "a" computer.... it's dozens or hundreds of computers... much like our new CPUs where more cores = better... the super computer is "more computers = better"....

The reasons for only rough estimations of the actual power is because they don't always fluidly "combine" as a clean total for each load... 100 computers at 10 teraflops doesn't necessarily give 1000 teraflops all the time... that would be the MAXIMUM....
 
4600 nodes with 6 Nvidia Volta GPUs per node. 4600 X 6 = 27600 Volta GPUs?

Don't know the price of this thing, but it can pay itself off in a week if they mine Bitcoins with it.
 
"Not wanting to be outdone, the US has been constructing what it believes will be the world's fastest supercomputer upon completion."

Is this a US vs China thing, or is because these Summit builders, as a separate entity actually need the power, and will benefit from the power.

There is *never* enough computing power available for scientists, that's just a given...
 
4600 nodes with 6 Nvidia Volta GPUs per node. 4600 X 6 = 27600 Volta GPUs?

Don't know the price of this thing, but it can pay itself off in a week if they mine Bitcoins with it.

Not really:)
Current ROI time for GPU in mining is about 6 month, for any number of GPUs: more GPUs = more profit = more money to return.
 
That
Not really:)
Current ROI time for GPU in mining is about 6 month, for any number of GPUs: more GPUs = more profit = more money to return.
That may be true, but GPUs based on Volta arent even available to the public, so we have no clue how quickly they would mine bitcoins. 1000,s of Volta gpus must be a pretty penny though, so I will go with 6+month ROI. Still not bad for a multi million dollar supercomputer.
 
That
Not really:)
Current ROI time for GPU in mining is about 6 month, for any number of GPUs: more GPUs = more profit = more money to return.
That may be true, but GPUs based on Volta arent even available to the public, so we have no clue how quickly they would mine bitcoins. 1000,s of Volta gpus must be a pretty penny though, so I will go with 6+month ROI. Still not bad for a multi million dollar supercomputer.

Maybe that's exactly what they'll be using the computer for. Under the excuse of "science" and "national security" they'll be mining ethereal.
 
Back