字幕列表 影片播放
I'm Gary Grider. I'm the division leader for High-Performance Computing at
Los Alamos National Lab. So, Los Alamos has been in the supercomputing business
for a long, long time: from the 40s before it was really called supercomputing.
We actually took the first problem to the first computer in the United
States: the ENIAC. Subsequently, we built our own variant of ENIAC
called MANIAC here at the Laboratory. We've been involved in a lot of firsts.
Probably one of the most important firsts is the invention of the Monte
Carlo Simulation Method and that was invented by Nicholas Metropolis
at Los Alamos. Almost every simulation done today uses Monte Carlo in some
way and so it's a really big deal. A recent first was the first machine to
achieve petascale, petaflop computing and that was Roadrunner. Roadrunner
was quite an interesting turning point for the community I think and it's led
us toward “If we can compute at that scale, can we compute an order of
magnitude, or two or three, more to explore things we didn't think we
could explore?” and that led to Trinity and Trinity was really designed to
solve this really large problem that needed to be resolved out to the point
where you needed the multi-petabytes of memory running on one job for many
months at a time to solve one problem. We have a request for proposal now
that's out to buy a machine to be delivered in 2020 and that machine
will be probably three to six, maybe eight, times larger than Trinity is some way.
And the reason we're buying a machine in 2016 to be delivered in 2020 is because
we're buying something that no one's invented yet, and it's going to take us
three or four years working with the chosen vendor. And so the whole cycle
is really quite long, right? It's four years to buy something and work with the vendor
to develop the technology, then you finally get it here and it takes a year to make it work
because it's never worked before, and then you get four years of productive
computing out of it and then it goes away and you do it again. At any one
time we have to have multiples of these going at once because it takes
nine years to do them.