U.S. Invests $258M in Supercomputing Race with China(wsj.com)
wsj.com
U.S. Invests $258M in Supercomputing Race with China
https://www.wsj.com/articles/u-s-invests-258-million-in-supercomputing-race-with-china-1497549600
70 コメント
AWS is a terrible match for super computing hardware. Sure, there are some embarrassingly parallel workloads that don't need fast interconnects, but mostly latency or memory becomes the issue.
Dirty little secret:
Much the computing on these machines isn't supercomputing and doesn't need fast interconnects. In fact, many (most?) workloads would be best done in the cloud. Except we already bought these machines. So there's tons of frameworks around meant to use these machines like they are AWS pizza boxes. If it's inefficient due to an architectural mismatch, that's fine. You're still using the machine that was paid for, and the only numbers that matter is utilization, and those metrics aren't even defined in terms of CPU utilization. If you reserve 10 nodes, and only run on one node, it's still reported that a 10 node job ran for 6 hours, not that a 10 node job job ran at 10% efficiency for 6 hours.
As far as why the accounting is this way, it's assumed that users will attempt to maximize their usage since allocations are defined. But even then, you can blow through your allocation and generally ask for more. But that's not necessarily what happens.
In summary: Because we bought these big computers, we have to use them, even if the workload isn't a good match.
Much the computing on these machines isn't supercomputing and doesn't need fast interconnects. In fact, many (most?) workloads would be best done in the cloud. Except we already bought these machines. So there's tons of frameworks around meant to use these machines like they are AWS pizza boxes. If it's inefficient due to an architectural mismatch, that's fine. You're still using the machine that was paid for, and the only numbers that matter is utilization, and those metrics aren't even defined in terms of CPU utilization. If you reserve 10 nodes, and only run on one node, it's still reported that a 10 node job ran for 6 hours, not that a 10 node job job ran at 10% efficiency for 6 hours.
As far as why the accounting is this way, it's assumed that users will attempt to maximize their usage since allocations are defined. But even then, you can blow through your allocation and generally ask for more. But that's not necessarily what happens.
In summary: Because we bought these big computers, we have to use them, even if the workload isn't a good match.
Amdahl's law, if you spend more on interconnects you only speed up a subset of your workloads. But the reverse is also true so buying more nodes and less interconnects only speeds up a subset of your workloads. The optimum is going to seem wasteful on average.
https://en.m.wikipedia.org/wiki/Amdahl%27s_law
https://en.m.wikipedia.org/wiki/Amdahl%27s_law
Gcp has it's tpus though...
That's like 1/5 the cost of one F35 Flying Turkey?
And a mere 2% the cost of the USS Gerald Ford (which would probably be destroyed within 24 hours of active war with China)
Your premise is comical. Any war that involves destroying a US carrier and killing thousands of soldiers in doing so, will move to a nuclear stand-off almost immediately thereafter. China and Russia have always understood that. For China, the ability to sink a US carrier only matters in instances of absolute war, where hundreds of millions of people will die. Being able to sink a carrier if you're a major military competitor to the US, is low on your list of actual concerns in the end (the threat is worth drastically more than the actual ability; if you use the ability and succeed, it's WW3). Carriers are for force projection, some intimidation, and fighting nations not named China or Russia.
Since we happen to be talking about sinking carriers and the possibility of nuclear war, I thought I'd mention a video that brings this scenario to life: https://www.youtube.com/watch?v=2VZ3LGfSMhA
It's a fictional BBC news segment covering a "serious incident" between NATO and Russian forces.
"Tensions between NATO and Russia have been growing over the past few years, and the continuing military build-up in the Baltic region is increasing concerns that a serious conflict might be a real possibility. This dramatization depicts one possible scenario. We hope that it shows how fragile our world is, and how quickly it can be changed by military and political decision-makers who do not speak for the tens of millions of people whose lives they may well destroy."
It's only tangentially related to the conversation, but it's such an excellent work of art that I can't help plugging it. The authors appear to be anonymous, which is interesting. If you were affected by the movie Threads or any other nuclear fiction, you might find this video similarly powerful.
It's a fictional BBC news segment covering a "serious incident" between NATO and Russian forces.
"Tensions between NATO and Russia have been growing over the past few years, and the continuing military build-up in the Baltic region is increasing concerns that a serious conflict might be a real possibility. This dramatization depicts one possible scenario. We hope that it shows how fragile our world is, and how quickly it can be changed by military and political decision-makers who do not speak for the tens of millions of people whose lives they may well destroy."
It's only tangentially related to the conversation, but it's such an excellent work of art that I can't help plugging it. The authors appear to be anonymous, which is interesting. If you were affected by the movie Threads or any other nuclear fiction, you might find this video similarly powerful.
I would like to see the details if your tactical analysis on that.
I'd recommend these as an interesting starting point ([2] has videos):
[1] http://infoproc.blogspot.com/2017/02/on-military-balance-of-...
[2] http://infoproc.blogspot.com/2016/09/balance-of-power-in-wes...
[1] http://infoproc.blogspot.com/2017/02/on-military-balance-of-...
[2] http://infoproc.blogspot.com/2016/09/balance-of-power-in-wes...
I read through those, including some of the referenced articles. My opinion is that the DF-21 is overhyped, as is the vulnerability of CVBGs to satellite detection. I'm on my phone right now, so I'll circle back to details later.
The best toys are not really widely known or battle tested. However, consider people have had a long time to prepare for aircraft carriers without a direct conflict between major powers and aircraft carriers are easily tracked in real time by satellites. Further, they might as well be sitting still in terms of a high mach (AKA 10+) missiles.
PS: Remember, we where doing optical recognition for missiles twenty years ago. So, reasonably close is going to be a hit most of the time.
PS: Remember, we where doing optical recognition for missiles twenty years ago. So, reasonably close is going to be a hit most of the time.
> The best toys are not really widely known or battle tested.
I know, I worked on some of them.
> However, consider people have had a long time to prepare for aircraft carriers without a direct conflict between major powers and aircraft carriers are easily tracked in real time by satellites.
As I mentioned in another reply, it isn't as easy as some people seem to think to find and track a CVBG by satellite. As I also mentioned, I'll go into more detail when I'm not on my phone.
> Further, they might as well be sitting still in terms of a high mach (AKA 10+) missiles.
I'm not aware of any hypersonic cruise missiles in that range, and those that approach it have relatively short ranges. I would need to look at this further.
> PS: Remember, we where doing optical recognition for missiles twenty years ago. So, reasonably close is going to be a hit most of the time.
Longer, I believe. Image recognition for terminal guidance is not foolproof, though. Plus, as with any weapons system the countermeasures groups have not been sitting still.
I know, I worked on some of them.
> However, consider people have had a long time to prepare for aircraft carriers without a direct conflict between major powers and aircraft carriers are easily tracked in real time by satellites.
As I mentioned in another reply, it isn't as easy as some people seem to think to find and track a CVBG by satellite. As I also mentioned, I'll go into more detail when I'm not on my phone.
> Further, they might as well be sitting still in terms of a high mach (AKA 10+) missiles.
I'm not aware of any hypersonic cruise missiles in that range, and those that approach it have relatively short ranges. I would need to look at this further.
> PS: Remember, we where doing optical recognition for missiles twenty years ago. So, reasonably close is going to be a hit most of the time.
Longer, I believe. Image recognition for terminal guidance is not foolproof, though. Plus, as with any weapons system the countermeasures groups have not been sitting still.
In terms of satellite tracking, even 10 meter resolution is enough to vastly reduce the search space. This can then be further refined though various means. Further, the search space is generally limited to how far they can travel from the last known location and they are not that fast.
But the real problem is more basic. They are very high value targets an can be taken out for vastly less money than they cost to produce.
But the real problem is more basic. They are very high value targets an can be taken out for vastly less money than they cost to produce.
http://www.zerohedge.com/news/2017-05-10/china-warns-washing...
Taken with a grain of salt: China might not actually be 'there' yet, but they are progressing toward that goal.
Taken with a grain of salt: China might not actually be 'there' yet, but they are progressing toward that goal.
I take everything from ZeroHedge with a block of salt...
Anyway, I would need to look into that. Chinese claims have a tendency to be blown out if proportion, particularly by the media of their neighbors.
Anyway, I would need to look into that. Chinese claims have a tendency to be blown out if proportion, particularly by the media of their neighbors.
Nuclear weapon. I'd give it 20 minutes tops.
If your tactical analysis involves nuclear weapons, there's no point in doing a tactical analysis, because most everything dies. Everyone knows this, which is why everyone is still building conventional weapons.
Why would China be working on conventional ship-to-ship missiles if they expected a nuclear war? They'd just use ICBMs / long-range missiles.
Why would China be working on conventional ship-to-ship missiles if they expected a nuclear war? They'd just use ICBMs / long-range missiles.
why are Americans so hell bent on having a war with China? last I heard China doesn't have military bases around the globe like the US does.
China likely fields the largest and most powerful military industrial complex in the world, or at least one of the largest. It makes sense to use China as a yardstick against which to measure the value of potential military expenditure.
I'm sure that when China does the same, they're not imagining the cost of war against, say, Australia or Mexico. they invoke the bogeyman of American military hegemony.
I'm sure that when China does the same, they're not imagining the cost of war against, say, Australia or Mexico. they invoke the bogeyman of American military hegemony.
No, it's 2 F35s. Unit cost will be a bit under 100mm each.
Does that price include oxygen? Not all of them do [0].
[0] https://www.theregister.co.uk/2017/06/13/f35_pilots_report_h...
[0] https://www.theregister.co.uk/2017/06/13/f35_pilots_report_h...
You'd like to think they'd investigate why some of their new planes cause hypoxia. Unfortunately, the scale of the problem makes it easier to try to understand why one variant (Marine Corps) ISN'T having the problem.
https://news.usni.org/2017/06/14/aviation-leaders-still-unsu...
https://news.usni.org/2017/06/14/aviation-leaders-still-unsu...
Unit costs can only be determined after enough runs permit you to take the spent budget and divide by the number of units built. What you're describing is a projection and it only holds if the full run is built. If it gets cut short, the unit costs are higher because of the non-recoverable costs.
Hahahha! Also roughly the price of just 2 Flying Raptors, which have the distinct honor of being nearly unfliable by even the most skilled pilots.
top500.org, which maintains the list of the 500 fatest supercomputers in the world also has more details on this [1]
> The PathForward program [...] address the four key challenges: parallelism, memory and storage, reliability and energy consumption
Energy consumption is becoming a very important challenge, as the article says, current petascale supercomputers consume "enough power to run a small town". Major efforts will need to be made to have Exascale supercomputers that don't consume the equivalent of 10^3 small towns.
[1] https://www.top500.org/news/doe-shells-out-258-million-to-si...
> The PathForward program [...] address the four key challenges: parallelism, memory and storage, reliability and energy consumption
Energy consumption is becoming a very important challenge, as the article says, current petascale supercomputers consume "enough power to run a small town". Major efforts will need to be made to have Exascale supercomputers that don't consume the equivalent of 10^3 small towns.
[1] https://www.top500.org/news/doe-shells-out-258-million-to-si...
Competition is good. For research, The Cold War with Russia and the resulting "Space Race" was good. In Israel, they spend a much larger share of their national income on R&D, largely to develop technologies to defend themselves.
This. I'd argue one of the biggest problems facing the United States right now is that there is a (falsely) perceived lack of competition / existential threats which is making us lazy and complacent.
This is just stupid. US basically sent out a signal to its competitor to demand them to invest more heavily on their own R&D capabilities- what Chinese universities, research institutions and companies are going to do with such news? They take it to the Chinese central government to ask for more R&D funding.
Think about it - what China can possibly lose in such a race? When Americans are trying to maintain its No.1 ranking in the field, Chinese are just trying to further modernize its industry which stands at less than $10k GDP per capita, you can probably beat the next generation Chinese supercomputer, but hey, you are pushing for the creation/improvement of the Chinese supercomputing community!
Think about it again - what China can possibly lose in such a race? America won't lose its overall dominance just because some other countries have faster supercomputers, America is losing its influence around the global when it stupidly using its borrowed resources to force its main competitor (also No.1/2 debt owner) for more R&D spending.
Think about it - what China can possibly lose in such a race? When Americans are trying to maintain its No.1 ranking in the field, Chinese are just trying to further modernize its industry which stands at less than $10k GDP per capita, you can probably beat the next generation Chinese supercomputer, but hey, you are pushing for the creation/improvement of the Chinese supercomputing community!
Think about it again - what China can possibly lose in such a race? America won't lose its overall dominance just because some other countries have faster supercomputers, America is losing its influence around the global when it stupidly using its borrowed resources to force its main competitor (also No.1/2 debt owner) for more R&D spending.
More R&D spending is usually good for humanity, no matter what country foots the bill, so I'm okay with this development.
Son, you're not making too much sense. Of all things to be mad about, tech investments isn't one of them.
If supercomputer is a race and the race is between the US & China, then such investment from the US federal government using borrowed money is _not_ doing any favour to the US - it might build a few faster supercomputer in the end, but in the long run, such a direct head on competition which fuels Chinese R&D spending in arguably one of the most cutting edge fields is damaging US interests.
Why Chinese are building supercomputers on their own? Because the US laws prohibit any US supercomputer export to China, including the export of core components that can be used to build such supercomputers.
In case you are not familiar with the matter - the above mentioned core components are not some fancy parts guarded by your NSA - we are talking about parts that can be freely purchased on ebay for $300-1000 each.
Why Chinese are building supercomputers on their own? Because the US laws prohibit any US supercomputer export to China, including the export of core components that can be used to build such supercomputers.
In case you are not familiar with the matter - the above mentioned core components are not some fancy parts guarded by your NSA - we are talking about parts that can be freely purchased on ebay for $300-1000 each.
This is chump change, we should have invested $2.58 billion or more. China isn't taking half measures in their state's science and technology investments. They realize we're in a second cold war, just one where the two combatants have more economic interdependence. They are making investments like the ones that helped us win the last war.
> They realize we're in a second cold war...
For God's sake, don't start another cold war. This is an established incumbent, who has monopolized world power for decades, feeling threatened by a new kid in town. Any progress China has made has been labeled as threat, cold war, or dismissed as useless, or blamed on China's stealing US secrets. As an instance, China bought many Intel chips to create supercomputers. Then the U.S. government banned Intel from selling high end chips to China[1]. China had to create homemade chips for world's fastest supercomputer[2].
[1] http://wccftech.com/us-government-bans-intel-nvidia-amd-chip... [2] http://www.pcworld.com/article/3086107/hardware/chinas-secre...
[1] http://wccftech.com/us-government-bans-intel-nvidia-amd-chip... [2] http://www.pcworld.com/article/3086107/hardware/chinas-secre...
I love this very naive view. China is 100% trying to beat and surpass the US as a superpower. They are not benevolent at all. Have a look at the ridiculous claims on the South China Sea for a start. I have spent lots of time in China. They still want payback for the humiliation they believe they endured in the 1800s. That is a one of the single most important driving factors in everything they do.
I don't know much about Chinese history. What humiliation did they endure in the 1800s?
I thought you would have linked to: https://en.wikipedia.org/wiki/Century_of_humiliation
Adding onto this, consider that China was, for a relatively long period in its development, the dominant power in its interactions. Interactions with the colonial West in the 1800s were a dramatic departure from long term history in a number of ways. At the end of the 1800s, the Sino Japanese War was also a major defeat for the Chinese. From a certain perspective, the current state of affairs could be viewed as something of a long-term return to normal.
"the current state of affairs could be viewed as something of a long-term return to normal"
CCP has its own "Make China Great Again" thing, officially known as "The great rejuvenation of the Chinese nation".
https://www.amazon.com/CHINESE-DREAM-GREAT-REJUVENATION-NATI...
CCP has its own "Make China Great Again" thing, officially known as "The great rejuvenation of the Chinese nation".
https://www.amazon.com/CHINESE-DREAM-GREAT-REJUVENATION-NATI...
You need to chill out and look at the future, not the past. If you dwell on the past, many countries have invaded many other countries. The British colonized and ruled the US for a long time. The US invaded Mexico and seized large portion of California and Texas. Even recently, the US invaded Yugoslavia, Iraq, Afghanistan, Syria, among others. China was the leading power for hundreds of years, way longer than the US or Britain. China rarely colonized other countries. The Chinese people look at the future, not the past.
The Chinese people are not the Chinese government. When someone is speaking about what "China" wants, it is the ruling party.
"I have spent lots of time in China. They still want payback for the humiliation they believe they endured in the 1800s."
"What I was trying to say is that my opinion is not informed based on what I have read in the press but based on the fact that I have spent time in the country and have read the local press, and spoken with the local people. The humiliation faced by the Chinese at the hand of the western powers is a common theme of books and movies and TV shows in China."
Not this someone.
"What I was trying to say is that my opinion is not informed based on what I have read in the press but based on the fact that I have spent time in the country and have read the local press, and spoken with the local people. The humiliation faced by the Chinese at the hand of the western powers is a common theme of books and movies and TV shows in China."
Not this someone.
When one is talking about Chinese territorial aggression and their desire to surpass the US as the geopolitical superpower, it is definitely not about the individual people.
Nothing the GGP said about the national Chinese interests was remotely controversial (perhaps subsequent comments are unrepresentative of individuals' thinking). Heck, Obama wanted to pass TPP mostly as a bulwark against Chinese ambitions.
Nothing the GGP said about the national Chinese interests was remotely controversial (perhaps subsequent comments are unrepresentative of individuals' thinking). Heck, Obama wanted to pass TPP mostly as a bulwark against Chinese ambitions.
It's not like the DoE are the only ones with expensive computers: The NSA has their underground lair.
There's a commercial need for expensive distributed processing now, with Google/AWS/Azure spending a lot on video cards. The ML craze may make government spending on supercomputers less important, especially if people prefer to distribute calculations on many interchangeable computers.
I wouldn't mind a few NSF or DoD grants for novel computing architectures, but I'm not sure I buy that having massive supercomputers for economic forecasting is "strategically important for nations".
There's a commercial need for expensive distributed processing now, with Google/AWS/Azure spending a lot on video cards. The ML craze may make government spending on supercomputers less important, especially if people prefer to distribute calculations on many interchangeable computers.
I wouldn't mind a few NSF or DoD grants for novel computing architectures, but I'm not sure I buy that having massive supercomputers for economic forecasting is "strategically important for nations".
these things get run in phases. phase 1 they insist on novel architectures, rethinking of basic premises, required involvement by academics.
by the end of phase 3 when they actually procure machines its mostly 'just give me one of what you're already shipping, and it better run MPI well'
doe exascale was supposed to be fundamentally different. because by the time you got there all the incremental improvements in power and latency management weren't nearly enough anymore...my guess is that they'll just make big gpu clusters and call it a day
its probably not that much money in the scheme of things, but the sad (or good) depending on your perspective is that HPC got so far out of the mainstream that when corporations finally got around to worrying about scaling, they just did their own thing. so I guess they can thank darpa/doe for infiniband? not really?
and they're still running fortran/mpi
by the end of phase 3 when they actually procure machines its mostly 'just give me one of what you're already shipping, and it better run MPI well'
doe exascale was supposed to be fundamentally different. because by the time you got there all the incremental improvements in power and latency management weren't nearly enough anymore...my guess is that they'll just make big gpu clusters and call it a day
its probably not that much money in the scheme of things, but the sad (or good) depending on your perspective is that HPC got so far out of the mainstream that when corporations finally got around to worrying about scaling, they just did their own thing. so I guess they can thank darpa/doe for infiniband? not really?
and they're still running fortran/mpi
Honestly, faster super computers aren't going to affect the balance of power. There is a ridiculous amount of computer power available already.
I'm not so sure about that.
Faster super computers allow you to do all kinds of things that would otherwise be quite hard. Think of supercomputers as weapons that you can fire very quietly, nobody even notices that you just tested some new (possibly nuclear) weapons system or that you designed some fighter jet, submarine or rocket.
Super computers are a game changer, not having them is as much as ceding territory, having them but not knowing how to use them has a similar effect. Just like computers were a game changer in the previous war so super computers will be a game changer in the next, with the added caveat that you might not even notice when it starts or ends other than that you suddenly find out that your counterparty runs the show because they cracked the codes you thought were safe.
Faster super computers allow you to do all kinds of things that would otherwise be quite hard. Think of supercomputers as weapons that you can fire very quietly, nobody even notices that you just tested some new (possibly nuclear) weapons system or that you designed some fighter jet, submarine or rocket.
Super computers are a game changer, not having them is as much as ceding territory, having them but not knowing how to use them has a similar effect. Just like computers were a game changer in the previous war so super computers will be a game changer in the next, with the added caveat that you might not even notice when it starts or ends other than that you suddenly find out that your counterparty runs the show because they cracked the codes you thought were safe.
I didn't say they weren't useful, but, as noted, we already have tons of processing power. Diminishing returns kick in at some point with any sort of investment like this. Tossing a few more super computers at labs isn't going to shake the world.
You can even get time on machines that aren't in the country. I'm sure Chinese experiments have run on US supercomputers at this point.
You can even get time on machines that aren't in the country. I'm sure Chinese experiments have run on US supercomputers at this point.
Some non-paywalled information at Anandtech: PathForward: US Dept. of Energy Awards $258M in Research Contracts To Develop Exascale Supercomputer Technology [0]
[0] http://www.anandtech.com/show/11547/us-dept-of-energy-awards...
[0] http://www.anandtech.com/show/11547/us-dept-of-energy-awards...
only 200M? Am I the only one think it is a small budget for supercomputing?
Here's the same content at AnandTech (w/o paywall):
http://www.anandtech.com/show/11547/us-dept-of-energy-awards...
http://www.anandtech.com/show/11547/us-dept-of-energy-awards...
Tweeting will be so fast, it is going to be great
It'll be a shame if they don't dump a chunk of that money on the Mill. However there may very well be some string attached unsuitable to Ivan Godard et al.
The Mill isn't really very well suited for supercomputing workloads. It tries to do well on poorly written C++ code (aka "general purpose code") that OoO machines have historically don well on. By contrast, VLIW machines are perfectly suitable for supercomputing, and indeed, I think that Itanium did well in the HPC world until it was discontinued.
Something like the Rex Computing Neo would be more suitable for this.
Something like the Rex Computing Neo would be more suitable for this.
I wonder if it being completely open source would have an impact on that.
Why is this in the DOE budget?
This is where the US' climate change budget is going [citation needed]
Guess this now needs reconfiguring http://worrydream.com/ClimateChange
Guess this now needs reconfiguring http://worrydream.com/ClimateChange
Pennies.
Does betteridge's law apply to conference talks?
http://sc14.supercomputing.org/schedule/event_detail-evid=pa...
They shoulda invested 256 million, woulda been more efficient.
I've spent a lot of time in country X. They all want Y.
Okay. That's the kind of crap that was in the 1800's, that's still in the 2010's, unfortunately.
Your words sound like something from a grown up bully jealous of their victim's achievements in life.
Okay. That's the kind of crap that was in the 1800's, that's still in the 2010's, unfortunately.
Your words sound like something from a grown up bully jealous of their victim's achievements in life.
> Your words sound like something from a grown up bully jealous
This crosses into incivility and personal attack, and those things are not allowed here. Please edit all that out of your comments when you post.
We detached this subthread from https://news.ycombinator.com/item?id=14565533 and marked it off-topic.
This crosses into incivility and personal attack, and those things are not allowed here. Please edit all that out of your comments when you post.
We detached this subthread from https://news.ycombinator.com/item?id=14565533 and marked it off-topic.
Our scientists look at what could be, look at their very interesting datasets, and then gnash their teeth for want of better tooling. People know this. Senior people from industry have screamed at senior government leadership that they are 20 years behind and falling faster. Nothing changes. Why?
Every time someone gets out from under their grad school payoff contract, they run to Google or Facebook.
Very few people are committing for the long term, and that's what's needed. Long-term commitment to steward a program. The people who remain often lack technical chops, have "certificates" in "cyber" something or other, but got through the Acquisitions school and now have authority to just pour money into blades, load Red Hat and mpi4py, and say "look, it's a supercomputer"!
They don't need to invest in iron. They need to invest in people. But they think they can just keep hiring new grads and expect magic.