Either Google or Amazon has the world's fastest supercomputer. The notion that China ever led that, was beligerently fraudulent (only good for selling sensationalism), it intentionally ignored that the nature of supercomputing has completely changed over the last 20 years.
> The notion that China ever led that, was beligerently fraudulent
The same could be said of the Soviet Union in the space, missile and other technological races. Two responses. One, it's a beneficial and not particularly risky lie. Two, the downsides to mistakenly thinking it's a lie are greater than those to mistakenly thinking it's the truth.
In what sense? The ranking is still based on LINPACK scores, and it's intended to represent the fastest machines for scientific computing.
We can argue about whether modern simulation codes look like LINPACK (they don't) but neither Google's nor Amazon's cloud is competitive for that type of workload. Even if the clouds are physically bigger, nobody at Google or Amazon is running really big numerical jobs of the sort that DOE does. Many are running smaller scale HPC and ML jobs, but it's expensive, and for latency-bound physics simulations (most of what we run at LLNL, ORNL, etc.), cloud networks will not perform at scale.
The question of whether China's machines ever performed well for big multi-physics applications is up in the air, but they were certainly beating us on LINPACK. The Chinese set out to show that they're serious about HPC, and I think they accomplished at least that. I wouldn't be surprised if you saw this race go back and forth a few times in the coming years.
As you say LINPACK is mostly useless as a benchmark.
But it's a wider problem than that. Interconnects expand the possible workloads, but that's always job specific. A single overclocked CPU would crush this for a huge range of tasks, but that in no way means it's not a super computer.
So, sure this thing is going to be useful, but Google still has vastly more compute and Google uses it for more economically valuable jobs.
The problem with dense linear algebra as a metric of supercomputer performance is that linear algebra tends to have relatively little communication compared to the computation. This means that the architecture you can use to really optimize linear algebra don't necessarily generalize well to different kinds of HPC code. China's architectures have definitely been accused of this in the past, but it looks like some of their newer stuff at least works fairly well on more HPC code.
I would be surprised if Google or Amazon had an existing capability that could match this machine on linpack. Stringing together FLOPS isn't the challenge. It's constructing an interconnect that can support the bandwidths needed to keep the compute units busy and actually hit anything close to what theoretical peak performance they are capable of while solving a specific problem. Google and Amazon have machines built for very different purposes.
What? Summit had been in development for years, and is ridiculously efficient for a machine with so much performance. It's a major milestone on the path to exascale computing.
How did this make it to the top 30... This "race" has been a global endeavour since the 50's, and no one keeps the title because that's not how progress works.
I don’t think we ever talk about them here. The rankings are fairly irrelevant to what most of us do, and the benchmarks used to derive them are too narrow anyways.
I mean I don’t really want to argue about it because it seems like a silly thing to argue about, but just because you aren’t interested doesn’t mean nobody else is. Here’s a thread from the November list: https://news.ycombinator.com/item?id=15707800
Google & Amazon also aren't reporting their computer fleet as well. However its likely Google is #1 on that list on raw compute available to R&D efforts.
I'm not so sure. It's normally about the interconnect with supercomputers, but modern datacenters have really pushed that boundary quite a bit. They still have higher latency in some regards, but they have a lot of reconfigurable hardware on the ethernet adapters that pushes down the latency in interesting ways.
I think he/she is talking about google TPU pods which can do 11.2 petaflops each. That's what they are offering google compute customers. Internally they probably have faster and larger pods by now as they usually publish about this stuff 6-12 months after they have moved on to something better.
It depends what you mean by a supercomputer. Having a lot of computers in data centers doesn't mean it's a supercomputer. Of course you can change the definition and call it # of cpus or compute "capacity" and such.
But seriously though. Why does the news media seem to always feel the need to push this narrative of nations being in competition with each other? It seems so 20th century.
Competition is great but when the narrative starts leaning towards "Us vs Them" and "We are at war" and countries start actively sabotaging each other, then we have a problem.
People steal technology because someone else’s is better and it’s cheaper than doing the R&D yourself. If the US just quietly made better computers and there was no ranking, I’m certain China and Russia would still be stealing IP.
If anything in this situation can be said to be "so 20th century," it's the notion that the bonds of trade and friendship transcend conflicts of interest, competition, and (god forbid) war.
The labs and supercomputer manufacturers also need this narrative to get continued funding from Congress. I spent many years in DOE supercomputing labs, and there was a heavy emphasis on competition whenever you wanted to knock loose a few hundred million dollars to buy a new computer.
>>This is one of the reasons why they're losing so much trust.
Sorry, but this is the actual bullshit narrative. Most media outlets, especially ones that Trump loves to rail against, are reporting record profits. That, to me, shows that people trust them more than ever.
The narrative sucks but nations being competitive with each other is a pretty powerful driving force for people.
I'm not sure about the U.S., but from my personal experience working in a multi-national industry, and also being East Asian myself, patriotism is a huge motivational factor, especially when it comes to beating "rival" countries. (Prime example: Korea vs Japan).
Yeah but it seems so artificial tho. What about doing things just for the sake of doing them and because they're intrinsically cool and interesting? All the nationalistic stuff is just made up and we've gotta let go of it eventually. Why not now?
I’m pretty sure that if nations wouldn’t compete we would be at peace and we would put all that useless military money into much more useful things. So no, I think this is not a powerful drive.
Form factors exist because they are local optima for non-computational concerns (ergonomics, cooling, electric, hardware maintenance, etc). Computers get better within the chosen form factors.
Calling it a "race" is really strange. They're not trying to build the first supercomputer, they're trying to build the fastest/most powerful supercomputer.
For instance, let's say China gets theirs out first, but the US releases one several months later that is substantially more powerful. Did China win the "race"? Or did the US?
I don't think calling it a race is all that strange. It's perhaps more strange to talk about winning the race or beating another country though, as that implies some finality/conclusion to the endeavor.
This can be looked at as a continuous race to see who can build the highest-throughput computer, where the competitors' positions in the race are decided based on a well-defined set of benchmarks. The periodic updates don't indicate who's won -- they're just a snapshot of what everyone's positions are in the race around the time of publication. So it's more apt to say one competitor is ahead of another in the race, but talking about winning or beating someone maybe gets more clicks.
Thanks, was looking for this info in both the article and the HN comments. It's not in the article, had to scroll halfway of the way down the comments to find it.
This "race" is starting to look like speed records for airplanes. There was a time when this was interesting technologically but now it seems more about the willingness to spend the money to get the record.
POWER9 CPUs! It's nice to see some competition in the area. If anything it at least has a better security / firmware story if you don't feel like running an embedded OS with backdoors in it https://www.intel.com/content/www/us/en/architecture-and-tec... that give attacker remote access
And it's almost affordable https://raptorcs.com/TALOSIILITE/ you might think it's a
joke but $1k does seem cheap compared to the prices for motherboards in their main product.
What kind of cores do machines of this caliber use?
I seem to recall that some years ago they used some Xeons and other "consumer-level" processors, just with special interconnects and maybe even using special versions, but they seemed like pretty close cousins to the regular processors of a desktop.
Is this still the case? I see in the Top500 list that the Sunway TaihuLight uses the Sunway SW26010 260C processors.
Are they fundamentally different than consumer-level processors? any special architectures or other interesting parts? or are they pretty similar, just built for massive interconnection and such?
Most of the time they care more about perf/watt than raw power. The problems that run on these have to be so embarrassingly parallel that individual node perf isn't that important. Sometimes those two line up well, but not always. For instance Blue Gene used more or less off the shelf PowerPC 4xx cores that were designed for embedded use cases. The neat stuff tends to be in the supporting chipset, particularly around the interconnect.
In this case it's using more or less off the shelf POWER9s and Nvidia GPUs.
57 comments
[ 2.7 ms ] story [ 136 ms ] threadEither Google or Amazon has the world's fastest supercomputer. The notion that China ever led that, was beligerently fraudulent (only good for selling sensationalism), it intentionally ignored that the nature of supercomputing has completely changed over the last 20 years.
The same could be said of the Soviet Union in the space, missile and other technological races. Two responses. One, it's a beneficial and not particularly risky lie. Two, the downsides to mistakenly thinking it's a lie are greater than those to mistakenly thinking it's the truth.
We can argue about whether modern simulation codes look like LINPACK (they don't) but neither Google's nor Amazon's cloud is competitive for that type of workload. Even if the clouds are physically bigger, nobody at Google or Amazon is running really big numerical jobs of the sort that DOE does. Many are running smaller scale HPC and ML jobs, but it's expensive, and for latency-bound physics simulations (most of what we run at LLNL, ORNL, etc.), cloud networks will not perform at scale.
The question of whether China's machines ever performed well for big multi-physics applications is up in the air, but they were certainly beating us on LINPACK. The Chinese set out to show that they're serious about HPC, and I think they accomplished at least that. I wouldn't be surprised if you saw this race go back and forth a few times in the coming years.
But it's a wider problem than that. Interconnects expand the possible workloads, but that's always job specific. A single overclocked CPU would crush this for a huge range of tasks, but that in no way means it's not a super computer.
So, sure this thing is going to be useful, but Google still has vastly more compute and Google uses it for more economically valuable jobs.
Wrong [0]
> This "race" has been a global endeavour since the 50's
From the article:
> But the United States retook the lead thanks to a machine, called Summit, built for the Oak Ridge National Laboratory in Tennessee.
When computers are used for nuclear weapons simulation, it's clearly not a global endeavor.
[0] https://www.top500.org/lists/2017/11/
The DOD stopped reporting many of their computers to the list though, so it’s anyone’s guess where those would land.
Edit: US weapons work that is. I would imagine China uses at least one of theirs for classified work.
But seriously though. Why does the news media seem to always feel the need to push this narrative of nations being in competition with each other? It seems so 20th century.
Early 20th century, I might add.
This is one of the reasons why they're losing so much trust.
Sorry, but this is the actual bullshit narrative. Most media outlets, especially ones that Trump loves to rail against, are reporting record profits. That, to me, shows that people trust them more than ever.
I'm not sure about the U.S., but from my personal experience working in a multi-national industry, and also being East Asian myself, patriotism is a huge motivational factor, especially when it comes to beating "rival" countries. (Prime example: Korea vs Japan).
This is how you get your budget cut. Competition forces tough decisions and turns long-term strategic aims into near-term threats.
Competition around destructive or hateful things is bad.
I'm OK with another space race, but not another nuclear arms race.
The space race was basically a PR veneer over part of the nuclear arms race.
A least some things never change.
For instance, let's say China gets theirs out first, but the US releases one several months later that is substantially more powerful. Did China win the "race"? Or did the US?
This can be looked at as a continuous race to see who can build the highest-throughput computer, where the competitors' positions in the race are decided based on a well-defined set of benchmarks. The periodic updates don't indicate who's won -- they're just a snapshot of what everyone's positions are in the race around the time of publication. So it's more apt to say one competitor is ahead of another in the race, but talking about winning or beating someone maybe gets more clicks.
Summit has significant higher peak performance at slightly higher power. It also has significantly fewer nodes, with higher performance per node.
Half as many nodes as Titan, way higher power efficiency, NVLink2 interconnect between CPU and GPU, ATS unified memory between all GPUs plus CPU, etc.
And it's almost affordable https://raptorcs.com/TALOSIILITE/ you might think it's a joke but $1k does seem cheap compared to the prices for motherboards in their main product.
I seem to recall that some years ago they used some Xeons and other "consumer-level" processors, just with special interconnects and maybe even using special versions, but they seemed like pretty close cousins to the regular processors of a desktop.
Is this still the case? I see in the Top500 list that the Sunway TaihuLight uses the Sunway SW26010 260C processors.
Are they fundamentally different than consumer-level processors? any special architectures or other interesting parts? or are they pretty similar, just built for massive interconnection and such?
In this case it's using more or less off the shelf POWER9s and Nvidia GPUs.