World’s Fastest Supercomputer Now Has Chinese Chip Technology

[jpiniero]

Homegrown is really interesting. Plus now since they bought AMD's R&D for cheap, they can speed up the process of eliminating any downsides to their own chips. You can bet they will copy anything useful they can get their hands on.

Bonus: China also now has more supercomputers in the Top 500 than the US.

 

[TheRyuu]

Each chip is supposedly as fast as an Apple A8 CPU[1], they've just got a lot of them. The interconnect technology is still from American companies.

The United States has only recently passed the National Strategic Computing Initiative after stagnating on the issue for some time so it's something to look forward to[1].

[1] http://www.nytimes.com/2016/06/21/technology/china-tops-list-of-fastest-computers-again.html

 

[Sonikku]

I'll believe it when it's bench marked. About all China is good at innovating is ways to steal designs of other businesses and re-appropriating them as their own.

 

[SarahKerrigan]

I'll believe it when it's bench marked. About all China is good at innovating is ways to steal designs of other businesses and re-appropriating them as their own.

It is benchmarked, running both Linpack (where it does rather well) and HPCG (where it does quite poorly - below both Tianhe-2 and K - likely due to an underperforming cache/memory hierarchy.)

 

[mxnerd]

To be on the list of Top500, of course it's benchmarked.

You have to submit the application and verified by Top500.org, which is a German company.

http://www.top500.org/project/call_for_participation/

 

[TheRyuu]

likely due to an underperforming cache/memory hierarchy.

It apparently uses an older memory technology so that's going to be a limiting factor.

 

[Phynaz]

It's ddr3-2100

 

[nenforcer]

It is believed it's on 28-nm process.

http://www.hpcwire.com/2016/06/19/china-125-petaflops-sunway/

It's still inconclusive if its .28nm according to that article. Just a few short years ago the best Chinese manufacturers were still at .90nm with planned transition to .65nm and .45nm. I would be surprised if was manufactured at .28nm domestically. I'm curious as to why this was left out of most all of the articles - probably for a reason.

 

[nenforcer]

us export rules are ignorant and counter-productive

This would have happened regardless of US export rules in order to reduce reliance on American tech.

 

[imported_ats]

Bonus: China also now has more supercomputers in the Top 500 than the US.

EH, not so much. China has had a bug up it butt about running HPL only pretty much every machine that exists. Things would change drastically if MS/Google/Amazon/FB/et al thought that running HPL was a good use of time or money for them. AKA, a decent number of the chinese machines in the Top 500 won't ever be used for HPC.

 

[beginner99]

This isn't due to export restriction but to Snowden leaks. Russia is doing something similar and rightly so. If Intel chips are back-doored, Russia and China would be fucked in case of war. It's probably easier to create your own chips than to validate Intel chips are 100% secure.
(Note: This has nothing to do with the other thread about backdoors in all Intel chips. Theory is NSA could have made intel to include some minor defects into the CPU design maybe also it'sccrypt extension and hence profit.)

It's still inconclusive if its .28nm according to that article. Just a few short years ago the best Chinese manufacturers were still at .90nm with planned transition to .65nm and .45nm. I would be surprised if was manufactured at .28nm domestically. I'm curious as to why this was left out of most all of the articles - probably for a reason.

Reason probably that they need a whole power plant to run that thing. If you have the money, are determined enough and economics/profitability don't matter, you can achieve a lot.

 

[DrMrLordX]

They allege that it's more efficient than tianhe-2, at least when running linpack.

 

[Headfoot]

From what I've heard, the really tricky part of designing an HPC is the interconnect / fabric heirarchy, not the CPU core. That would seem to apply here

 

[Essence_of_War]

From what I've heard, the really tricky part of designing an HPC is the interconnect / fabric heirarchy, not the CPU core. That would seem to apply here

And *surprise surprise*, lik tianhe2, this uses a tree rather than mesh architecture for communication between nodes. Much cheaper. Much easier to engineer. Much harder to do Non-linpack work on. I'd much rather work on Titan or sequoia...

 

[superstition]

This isn't due to export restriction but to Snowden leaks. Russia is doing something similar and rightly so. If Intel chips are back-doored, Russia and China would be fucked in case of war. It's probably easier to create your own chips than to validate Intel chips are 100% secure.

Any nation's leadership that doesn't assume backdoors is so naive that it's practically a fantasy leadership — unless we're talking about two-bit countries like Togo or something.

Espionage isn't a new thing.

 

[nenforcer]

Yeah these machines seem specifically tailored just to perform extremely well in the LINPACK benchmark.

 

[daxzy]

Yeah these machines seem specifically tailored just to perform extremely well in the LINPACK benchmark.

Yea, not shocking at all. People should stop using that top500 crap and replacement it with something that has more realworld stuff like SpecINT or SpecFP.

Here's HPCG benchmark (more real-world) for supercomputers. It would probably be rank 4 here (based on 0.003% of HLP figure given). In real-world applications, it'd probably be much lower.

http://www.hpcg-benchmark.org/custom/index.html?lid=155&slid=282

 

[soresu]

As for process, the previous gen Shenwei chip (Shenwei Sw-3 / SW1600) was released in 2010 at 65nm with 16 cores. Chances are pretty fair that they have gone 2+ nodes beyond 65nm in six years.

 

[AnitaPeterson]

So... in five years, we could see a market dominated by cheap Chinese CPUs in the home computing/DIY and server segments?

 

[DrMrLordX]

Doubt it. These CPUs on display are likely unsuitable for home/DIY usage.

 

[aigomorla]

So... in five years, we could see a market dominated by cheap Chinese CPUs in the home computing/DIY and server segments?

nope...

in 5 yrs i expect massive increases in IOP on ARM cpu's and more servers to slowly migrate over to ARM arch, over x86 or this cpu.

 

[ericlp]

1) it's no secret that Chinese have a lot more engineers than we do...
2) This server is all Chinese, even the backbone bus for chip communication is Chinese tech, they built that long ago on the past 2-3 servers running.
3) The handwriting was on the wall 10 years ago. It was only a matter of time.

As for migrating away from x86 chips, is what china just did.

4) china isn't just gonna sit around and wait for America to catch up! They are probably already working on the next gen that will be 10 times faster.

I could really give two shits who pushes the tech envelope. My hat comes off as I am grateful someone is making huge strides. China has been working on this for the last 20 years. Congrats! I can only hope that it lights a fire under US markets to compete and try to catch up. America has done a shitty job and just waits around. The tech here is lacking. Nothing new (radically) changed in the USA... we kinda just sat on our hands as Intel fires 15000 American Workers and applies for 15000 work visa's. What does that tell you??????

I say screw America, hire your own people to work.... I'm boycotting Intel just for that alone. Next chips will be from AMD, unless china can get me one!

 

[soresu]

Also bear in mind the 64 core custom m-arch ARMv8 chip called Mars, announced by chinese firm Phytium last year at HotChips.
That is a large chip (600mm2+ at 28nm) but it seems to be a first gen design, and they are working on increasing both areal, and power efficiency for second gen designs already.
So they certainly arent limited to MIPS Loongson or Alpha derived Shenwei even now, let alone in the future.
A quick scan on SMIC seems to indicate China is still at least 2-4 years from a finfet semi node though.

 

[ReignQuake]

Surely earning international money from processors makes more sense than this. They'll do their nuclear research anyway. Look how many Supercomputers they have. Sell to them before they no longer need us. These things will probably power their robot factories.

We borrowed €10 Billion for a nuclear power station. Chinese money is pretty good.

 

[Qwertilot]

Very sensible, but governments don't always think that logically about such matters Thing is, once they decide they don't want to rely on us for anything they won't have to in a decade or two.

We're talking about a huge chunk of the worlds population, with real resources/good education levels in places, and one which has always been capable - all the way back to the great wall - of really organised concerted pushes to achieve things.