Japan takes top spot on supercomputer list with ARM-powered machine

midian182

Posts: 9,658   +121
Staff member
What just happened? The bi-annual TOP500 Supercomputer list usually sees China and the US battling for the top spot, but not anymore. In the latest, 55th edition of the rankings, Japan has taken the number one position for the first time since 2011. It’s also the first instance of an ARM-based computer beating all others.

Japan’s system, Fugaku, which is installed in Kobe and developed by Fujitsu and the government-sponsored Riken institute, managed a High Performance Linpack (HPL) result of 415.5 petaflops. That makes it around 2.8 times faster than the now second-place Summit from IBM. The system also came first in the High-Performance Conjugate Gradient (HPCG) Benchmark with a record 13.4 HPCG-petaflops.

It took six years to develop Fugaku, which uses Fujitsu’s 48-core A64FX system-on-chip. It’s the first time an ARM-based processor has grabbed the top position—there are only four of these systems in the TOP500, three of which use the same Fujitsu processor. The x86 architecture is used in 481 systems, and 469 of these are from Intel.

Thanks to the addition of Fugaku, the aggregate list performance has jumped from 1.65 exaflops six months ago to 2.23 exaflops. The latest edition sees the lowest number of new systems (51) since the TOP500 began in 1993, possibly a result of Covid-19-related delays.

Back in 2017, China overtook the US when it came to the number of supercomputers on the list. Back then, it had 202 entries, compared to America’s 143. Now, China has 226, while the US has fallen to 114. Japan is third with 30, followed by France (18), and Germany (16).

Despite being second to China in terms of overall supercomputer numbers, the US is still number one in aggregate list performance, boasting 644 petaflops to its rival’s 565 petaflops.

Fugaku has been used for Covid-19 research, and it’s hoped the $1.2 billion machine will help identify treatments for the virus when it goes into full operation next year. It will also be used to model the impact of earthquakes and tsunamis and map out escape routes.

Permalink to story.

 
With this, plus Apple jumping to ARM, future starts looking grim for Intel and AMD.
Hardly, if you look at the numbers this ARM based system is 2.8 times faster which is impressive, until you realize it also consumes 2.8 times the power and has 3 times the number of cores. So by brute force it surpasses the now 2nd place system. If it was able to accomplish these results with less cores and power consumption by percentage it would be a truly impressive accomplishment.
 
Hardly, if you look at the numbers this ARM based system is 2.8 times faster which is impressive, until you realize it also consumes 2.8 times the power and has 3 times the number of cores. So by brute force it surpasses the now 2nd place system. If it was able to accomplish these results with less cores and power consumption by percentage it would be a truly impressive accomplishment.
If you look at where CPU technology goes, it is toward hyper-threading, because single-core performance is a dead end. ARM strength is in better scalability and less power consumption per core. Those are the key to features for the future CPU development.

The fact that we already have mobile-phone CPU-s in the top tier that can compete with desktop CPU-s speaks for itself. That, plus platform unification is a great recipe for changing the market. We are seeing the first major steps in that direction.
 
If you look at where CPU technology goes, it is toward hyper-threading, because single-core performance is a dead end. ARM strength is in better scalability and less power consumption per core. Those are the key to features for the future CPU development.

The fact that we already have mobile-phone CPU-s in the top tier that can compete with desktop CPU-s speaks for itself. That, plus platform unification is a great recipe for changing the market. We are seeing the first major steps in that direction.
Less power consumption per core is a great strategy in the mobile division where leveraging more cores when required or turning ones others when not to save power is great. However in a supercomputer where you would be using every available core at its full potential and rarely leveraging the power saving ability makes it almost irrelevant.

Again, although this new supercomputer is 2.8 times faster it achieved this with the use 3 times the number of cores, slower cores that individually use less power, but total power consumption is still roughly 2.8 times that of the existing 2nd place device. The efficiency of these two devices are nearly the same in other words.

Can you refrence what mobile-phone CPUs your saying compete with desktop CPUs?
 
...
Can you refrence what mobile-phone CPUs your saying compete with desktop CPUs?

I've seen Apple fans get a geekbench score on their ipad that makes them believe that they can edit video on it better than a 28 core Xeon ... "well the score says ...."

Every time I've seen people make such comparisons, its usually from an Apple fan that doesn't actually understand hardware.

So in that light, I too would like to see this reference to a phone having the same computing power as a desktop ... because so far the claims I've seen in this area are nonsense, but I'm willing to be proven wrong if someone can provide a solid and tight comparison.
 
Hardly, if you look at the numbers this ARM based system is 2.8 times faster which is impressive, until you realize it also consumes 2.8 times the power and has 3 times the number of cores. So by brute force it surpasses the now 2nd place system. If it was able to accomplish these results with less cores and power consumption by percentage it would be a truly impressive accomplishment.
It is worth noting that it is 2.8 times faster in the Linpack benchmark, but 4.8 times faster in the newer HPCG benchmark. https://www.top500.org/lists/hpcg/list/2020/06/

Apparently many common supercomputing tasks are better represented by the HPCG benchmark, which stresses the interconnect much more than Linpack. That is why in the standard top500 list, so many web servers with slower Ethernet connections have appeared, but these are not systems doing typical supercomputing tasks.
 
With this, plus Apple jumping to ARM, future starts looking grim for Intel and AMD.
That isn't likely. A CPU is serial meaning it can only clear a calculation by completing the calc on any individual core or thread, ARM has a disadvantage of being low power, meaning it can never equate to a higher ipc than a x86 based desktop processor, sure you can pack in lower frequency cores upping your overall level of compute, but your individual calculations will never finish faster in fact I would seriously say a quantum processor is a major leap in performance because even though it is limited it has the ability to be parallel in nature.

Even then what will you do when optronics start to overtake the market?Apple has a serious issue on their hands here single threaded performance will tank, the one good thing about Macs is now flushed down the toilet.
 
Hardly, if you look at the numbers this ARM based system is 2.8 times faster which is impressive, until you realize it also consumes 2.8 times the power and has 3 times the number of cores. So by brute force it surpasses the now 2nd place system. If it was able to accomplish these results with less cores and power consumption by percentage it would be a truly impressive accomplishment.
Basically yeah they pulled an AMD. If you can't beat them on ipc just beat them on cost and cores and fix it in the future, but as I stated a 1.2ghz core will never finish a calc faster than a 5ghz one and anything single threaded will suffer massively.
 
Less power consumption per core is a great strategy in the mobile division where leveraging more cores when required or turning ones others when not to save power is great. However in a supercomputer where you would be using every available core at its full potential and rarely leveraging the power saving ability makes it almost irrelevant.

Again, although this new supercomputer is 2.8 times faster it achieved this with the use 3 times the number of cores, slower cores that individually use less power, but total power consumption is still roughly 2.8 times that of the existing 2nd place device. The efficiency of these two devices are nearly the same in other words.

Can you reference what mobile-phone CPUs your saying compete with desktop CPUs?
But you know that the Fujitsu version of the ARM CPU can and will be improved on mightily compared to the ARMs available when this project started. We will not say farewell to arms, either in super computers or otherwise. So look for improved flops with lower power consumption, maybe in a couple of years. In stating this, I am assuming that Fujitsu uses TSMC as a chip fab, and TSMC's 5nm process is born to run.
 
Doesn't look too rosy for the US in the world of supercomputers, either.

Frontier will be ~3x more powerful than this ARM supercomputer. ETA 2021.

And it will be powered by AMD. I do think that RISC based processors will become more popular as we continue to find new ways to enable ease of programming with them, but as for it painting a "grim" picture for AMD and Intel, that's still quite a way out.

Also Note AMD is rumoured to use an "enhanced" version TSMCs 5nm node for Zen4 - which should arrive in late 2021.
 
Last edited:
I've seen Apple fans get a geekbench score on their ipad that makes them believe that they can edit video on it better than a 28 core Xeon ... "well the score says ...."

Every time I've seen people make such comparisons, its usually from an Apple fan that doesn't actually understand hardware.

So in that light, I too would like to see this reference to a phone having the same computing power as a desktop ... because so far the claims I've seen in this area are nonsense, but I'm willing to be proven wrong if someone can provide a solid and tight comparison.

of course someone who doesn't agree with you, must be stupid, that is just the way it is. for the record, no one, apple fan or not, thinks a Geekbench score of 4600 can do video edits anywhere near in the range of a 28 core Xeon based system, and you know that, so you shouldn't say it, snarky or not.

A whole lot of "apple fans" actually do understand hardware, sorry for the clarification, I know it hurts your world view. You do know there are desktops and there are DESKTOPS, right? To date, no one has built a phone/tablet anywhere the equivalent of a top end 6-core i7 or i9, AMD threadripper, or Xeon, but there are a whole lot of lower end desktops to which tablets do have the raw speed to be comparable. And that is the only thing alleged so far. and of course, even apple fans are aware of thermal limits on small form factor devices. But putting those same chips, let alone custom designs, in a desktop or laptop, with adequate cooling, gets around those limitations.

So your big mistake is thinking that anyone, including those gawdawful Apple fans, implied ever that a phone/tablet can compete with the higher end desktops available, any assertions made were simply that speeds in the ARM segment, even limited for small form factors and thermal limits, are approaching 4-core i5 and i7 class CPUs, not that they are a direct replacement or could be used in that capacity


 
of course someone who doesn't agree with you, must be stupid, that is just the way it is. for the record, no one, apple fan or not, thinks a Geekbench score of 4600 can do video edits anywhere near in the range of a 28 core Xeon based system, and you know that, so you shouldn't say it, snarky or not.

A whole lot of "apple fans" actually do understand hardware, sorry for the clarification, I know it hurts your world view. You do know there are desktops and there are DESKTOPS, right? To date, no one has built a phone/tablet anywhere the equivalent of a top end 6-core i7 or i9, AMD threadripper, or Xeon, but there are a whole lot of lower end desktops to which tablets do have the raw speed to be comparable. And that is the only thing alleged so far. and of course, even apple fans are aware of thermal limits on small form factor devices. But putting those same chips, let alone custom designs, in a desktop or laptop, with adequate cooling, gets around those limitations.

So your big mistake is thinking that anyone, including those gawdawful Apple fans, implied ever that a phone/tablet can compete with the higher end desktops available, any assertions made were simply that speeds in the ARM segment, even limited for small form factors and thermal limits, are approaching 4-core i5 and i7 class CPUs, not that they are a direct replacement or could be used in that capacity

I never implied anyone was stupid. I said that when I have seen people claiming that phone processors can outperform desktop processors (which I have encountered more than once), it has been because they are using a "benchmark" score that is not giving remotely an apples to apples comparison (geekbench is unreliable as hell to begin with). If they can't realize the error here, then it is what it is. it doesn't mean anyone is stupid. Its you who implied that of my post, not me.

I also never said apple fans are "gawdawful", again that was your implication; I don't think that at all, but by the way you completely projected false garbage onto my post and then responded as though I had said those things that you projected, you are beginning to highly compel me to change my mind. Apple fans tend to make themselves look bad ... hint, hint.

You are also stating that I am "making a big mistake" by misreading people when they claim "phones processors are as fast as desktop" ones ... how the hell do you misread that exact text? I was clearly referring to my anecdotal real life experience -- and you were not there.

Wow ... triggered much? :) Oh Wait! It was you, wasn't it!? ... ;)

Like I said, I'd love to be proven wrong, I am just waiting to see someone present a solid and tight comparison. Can you provide that?
 
Last edited:
But you know that the Fujitsu version of the ARM CPU can and will be improved on mightily compared to the ARMs available when this project started. We will not say farewell to arms, either in super computers or otherwise. So look for improved flops with lower power consumption, maybe in a couple of years. In stating this, I am assuming that Fujitsu uses TSMC as a chip fab, and TSMC's 5nm process is born to run.
I do not doubt the future potential of the ARM platform for use in the supercomputer sector, this Fujitsu monster is only the beginning and holds promise.
 
"It's also the first instance of an ARM-based computer beating all others."

Umm, I'm pretty sure that most of the Chinese supercomputers are also ARM-based so I believe this to be false.
 
"It's also the first instance of an ARM-based computer beating all others."

Umm, I'm pretty sure that most of the Chinese supercomputers are also ARM-based so I believe this to be false.
Ummm, I’m pretty sure the Chinese chips are RISC not ARM, to be ARM, they would have to use the proprietary architecture and instruction set and be licensed. But your overall point is made. RISC beating CISC. And that is before all the technical arguments which trivialize that statement and are very far over my head
 
Ummm, I’m pretty sure the Chinese chips are RISC not ARM, to be ARM, they would have to use the proprietary architecture and instruction set and be licensed. But your overall point is made. RISC beating CISC. And that is before all the technical arguments which trivialize that statement and are very far over my head
Strange, I distinctly remember reading that they were ARM-based (because ARM was open-source at the time) but I could have remembered it wrong or the source could have been incorrect. I don't remember where I read it so I'll stand corrected. Thanks for clearing that up.
 
Back