ARM powers the fastest Supercomputer.

[ArunMKumar]

Fugaku Supercomputers (https://en.wikipedia.org/wiki/Fugaku_(supercomputer)) has achieved 537 petaflops (200 petaflops by the previous Rank #1) of performance.

This hasn't created much discussion anywhere since it was announced, but will be soon released in the top 500 list (June 1st week)

 

here is the tweet: https://twitter.com/ProfMatsuoka/status/1261194036276154368?s=20

 

Quote

"We just announced the Fugaku config at a press briefing. The whole system is 158,976 A64FX nodes, and running at 2.2Ghz peak performances for 64bitFP/32bitFP/16bitFP/8bitINT are 537Peta/1.07Exa/2.15Exa/4.30Exa (FL)ops respectively, as well as the total memory BW 163 PetaByte/s.

This, if and when it trips down to the data centre has the ability to integrate the mobile and the server architecture, which the x86 hasn't been able to achieve till now.

scalable mobile games anyone?

wanted to see what opinions the forum has on it.

 

[LAwLz]

I wonder what the benefits of it being ARM based will be.

Obviously this computer is extremely powerful, but it would be interesting to know about efficiency and performance compared to what x86 could achieve with the same resources.

 

The performance per watt of the prototype, while impressive, is pretty much the same as what other computers with Xeon and POWER9 processors have achieved before.

[Sauron]

12 minutes ago, ArunMKumar said:

This, if and when it trips down to the data centre has the ability to integrate the mobile and the server architecture

Uhm... not really. I doubt these have much to do with the SoC you'd find in your average phone aside from the ISA.

13 minutes ago, ArunMKumar said:

which the x86 hasn't been able to achieve till now.

We had x86 phones up until recently. They still had nothing to do with server hardware.

14 minutes ago, ArunMKumar said:

scalable mobile games anyone?

No... and it's not like anyone is making games for megaclusters like this one...

[GDRRiley]

meh given there are at least 3 exscale supercomputers coming online soon that will break the barrier with some in the us by 2022 set to reach 1.5-2

[3rrant]

20 minutes ago, ArunMKumar said:

Fugaku Supercomputers (https://en.wikipedia.org/wiki/Fugaku_(supercomputer)) has achieved 537 petaflops (200 petaflops by the previous Rank #1) of performance.

This hasn't created much discussion anywhere since it was announced, but will be soon released in the top 500 list (June 1st week)

 

here is the tweet: https://twitter.com/ProfMatsuoka/status/1261194036276154368?s=20

 

This, if and when it trips down to the data centre has the ability to integrate the mobile and the server architecture, which the x86 hasn't been able to achieve till now.

scalable mobile games anyone?

wanted to see what opinions the forum has on it.

 

They announced the config but the system isn't up and running right now, isn't it?

By the 2021 which is the planned release there will be more than one hexascale x86 supercomputers (Aurora, Frontier.. ) if they all respect their release dates.

Impressive still, and probably will be at the top or near in the Green500 list but unlikely in the Top500

[ArunMKumar]

5 minutes ago, Sauron said:

Uhm... not really. I doubt these have much to do with the SoC you'd find in your average phone aside from the ISA.

We had x86 phones up until recently. They still had nothing to do with server hardware.

No... and it's not like anyone is making games for megaclusters like this one...

not yet, but the technology I believe will surely trio. down to the SoCs, not in the same size or performance.  x86 had its power efficiency issues in a mobile environment, which ARM till now has been really good at.

The software eco system around ARM has been developed considering low power usage too.

 

Gaming was a bad example.. TBH.. it was to meet the community guidelines. I could not think of something on the fly. This is too early a stage to suggest a commercial use-case.

[ArunMKumar]

2 minutes ago, 3rrant said:

They announced the config but the system isn't up and running right now, isn't it?

By the 2021 which is the planned release there will be more than one hexascale x86 supercomputers (Aurora, Frontier.. ) if they all respect their release dates.

Impressive still, and probably will be at the top or near in the Green500 list but unlikely in the Top500

From what I know, It will be announced this year. 

Others will surely try to catch up, HPC competition is fun, we know we would never have such computers at our desk but good to look at those numbers and admire them

[Sauron]

5 minutes ago, ArunMKumar said:

not yet, but the technology I believe will surely trio. down to the SoCs

How? What "technology" here is applicable to phones in your opinion?

[Jumballi]

The number of cpu used in it is telling, also we broke exoflop a while back.

[Jumballi]

1 minute ago, Sauron said:

How? What "technology" here is applicable to phones in your opinion?

the cloud

[ArunMKumar]

4 minutes ago, Sauron said:

How? What "technology" here is applicable to phones in your opinion?

The updated vector processing units have been worked upon for almost a decade. This can possibly come down to the integrated Mali GPU on the sone of the SoC. Adreno being the more prevalent with the snapdragon chips. 

Not anytime soon though. 

 

we can also see more software revisions for the ARM builds of enterprise software in the time to come if this takes off.

[GOTSpectrum]

1 hour ago, Jumballi said:

also we broke exoflop a while back.

I don't think that is quite right, unless I have missed out on some massive piece of news recently. You may be referring to The Folding@Home network surpassing the exaflop barrier but that isn't considered a supercomputer, but instead a distributed computing network. 

 

P.S. Come fold with us!  

[hishnash]

12 hours ago, LAwLz said:

I wonder what the benefits of it being ARM based will be.

The main benefit will be that the ARM cpus used were will be ultra-custom parts. if you want to do that with x86 you need to get AMD or Intel to do that work. 

 

At an instruction sets level the main difference is about how you optimise, RISC architectures assume you have a very good compiler optimisation, the aim is to have a small number of general purpose instructions but that each instruction runs very fast (almost all run within one cycle). So to optimise your compiler needs to work out what the best sequence of instruction ist. 

in x86 you have lots (and losts) of instructions and many of them take a very long time, to optimise x86 code much of the time your aim is to reduce the number of instructions you send to the cpu (aka you try to use things like AVX) but remember even through it is one instruction it might still take a long time to run on an x86 system, and will use a LOT of power, using AVX and lead the the cpu dropping well below its base clock speed.

In sort is is easy for a human to write optimised x86 since they can focus on finding the best single instruction. However it is harder for a compiler to turn generic code into fast x86. 

 

 

[BuckGup]

13 hours ago, LAwLz said:

I wonder what the benefits of it being ARM based will be.

Obviously this computer is extremely powerful, but it would be interesting to know about efficiency and performance compared to what x86 could achieve with the same resources.

 

The performance per watt of the prototype, while impressive, is pretty much the same as what other computers with Xeon and POWER9 processors have achieved before.

Often they can design the ARM chip around the work load so it’s essentially an ASIC. This is also the fasest public computer. There’s a couple faster computers that are private though that also utilize a similar ARM CPU

[Windows9]

[porina]

Writeup on it at Anandtech: https://www.anandtech.com/show/15869/new-1-supercomputer-fujitsus-fugaku-and-a64fx-take-arm-to-the-top-with-415-petaflops

 

I'd really love to try one of these CPUs, if I can get software rewritten for it. This sounds like the best of Skylake-X and Knights Landing put together. It not only has a load of FP64 potential, but also the bandwidth to feed it data. That's the biggest limitation in any x86 implementation so far. 4 groups of 12 compute cores (+1 assistant core), each fed with 256GB/s of bandwidth. About 21.3 GB/s per compute core. In x86 world, best case you're likely to see is something like a 7800X (6 cores, quad channel) which with ram at 3200 quad channel, works out about 12.5 GB/s per core. Anything Zen is far worse than that, even before considering the need to fragment data to CCX and and limited internal bandwidth.

[Results45]

50 minutes ago, porina said:

Writeup on it at Anandtech: https://www.anandtech.com/show/15869/new-1-supercomputer-fujitsus-fugaku-and-a64fx-take-arm-to-the-top-with-415-petaflops

 

 

That is some futuristic compute performance and bandwidth:

Quote

A64FX also supports quantized models with lower precision, which is where we get into some fun numbers for Fugaku:

• FP64: 0.54 ExaFLOPs
• FP32: 1.07 ExaOPs
• FP16: 2.15 ExaOPs
• INT8: 4.30 ExaOPs

Due to the design of the A64FX, it also allows for a total memory bandwith of 163 PetaBytes per second.

.