How China Built an Exascale Supercomputer Out of Old 14nm Tech (nextplatform.com) 29
Slashdot reader katydid77 shares a report from the supercomputing site The Next Platform:
If you need any proof that it doesn't take the most advanced chip manufacturing processes to create an exascale-class supercomputer, you need look no further than the Sunway "OceanLight" system housed at the National Supercomputing Center in Wuxi, China. Some of the architectural details of the OceanLight supercomputer came to our attention as part of a paper published by Alibaba Group, Tsinghua University, DAMO Academy, Zhejiang Lab, and Beijing Academy of Artificial Intelligence, which is running a pretrained machine learning model called BaGuaLu, across more than 37 million cores and 14.5 trillion parameters (presumably with FP32 single precision), and has the capability to scale to 174 trillion parameters (and approaching what is called "brain-scale" where the number of parameters starts approaching the number of synapses in the human brain)....
Add it all up, and the 105 cabinet system tested on the BaGuaLu training model, with its 107,250 SW26010-Pro processors, had a peak theoretical performance of 1.51 exaflops. We like base 2 numbers and think that the OceanLight system probably scales to 160 cabinets, which would be 163,840 nodes and just under 2.3 exaflops of peak FP64 and FP32 performance. If it is only 120 cabinets (also a base 2 number), OceanLight will come in at 1.72 exaflops peak. But these rack scales are, once again, just hunches. If the 160 cabinet scale is the maximum for OceanLight, then China could best the performance of the 1.5 exaflops "Frontier" supercomputer being tuned up at Oak Ridge National Laboratories today and also extend beyond the peak theoretical performance of the 2 exaflops "Aurora" supercomputer coming to Argonne National Laboratory later this year — and maybe even further than the "El Capitan" supercomputer going into Lawrence Livermore National Laboratory in 2023 and expected to be around 2.2 exaflops to 2.3 exaflops according to the scuttlebutt.
We would love to see the thermals and costs of OceanLight. The SW26010-Pro chip could burn very hot, to be sure, and run up the electric bill for power and cooling, but if SMIC [China's largest foundry] can get good yield on 14 nanometer processes, the chip could be a lot less expensive to make than, say, a massive GPU accelerator from Nvidia, AMD, or Intel. (It's hard to say.) Regardless, having indigenous parts matters more than power efficiency for China right now, and into its future, and we said as much last summer when contemplating China's long road to IT independence. Imagine what China can do with a shrink to 7 nanometer processes when SMIC delivers them — apparently not even using extreme ultraviolet (EUV) light — many years hence....
The bottom line is that the National Research Center of Parallel Computer Engineering and Technology (known as NRCPC), working with SMIC, has had an exascale machine in the field for a year already. (There are two, in fact.) Can the United States say that right now? No it can't.
Add it all up, and the 105 cabinet system tested on the BaGuaLu training model, with its 107,250 SW26010-Pro processors, had a peak theoretical performance of 1.51 exaflops. We like base 2 numbers and think that the OceanLight system probably scales to 160 cabinets, which would be 163,840 nodes and just under 2.3 exaflops of peak FP64 and FP32 performance. If it is only 120 cabinets (also a base 2 number), OceanLight will come in at 1.72 exaflops peak. But these rack scales are, once again, just hunches. If the 160 cabinet scale is the maximum for OceanLight, then China could best the performance of the 1.5 exaflops "Frontier" supercomputer being tuned up at Oak Ridge National Laboratories today and also extend beyond the peak theoretical performance of the 2 exaflops "Aurora" supercomputer coming to Argonne National Laboratory later this year — and maybe even further than the "El Capitan" supercomputer going into Lawrence Livermore National Laboratory in 2023 and expected to be around 2.2 exaflops to 2.3 exaflops according to the scuttlebutt.
We would love to see the thermals and costs of OceanLight. The SW26010-Pro chip could burn very hot, to be sure, and run up the electric bill for power and cooling, but if SMIC [China's largest foundry] can get good yield on 14 nanometer processes, the chip could be a lot less expensive to make than, say, a massive GPU accelerator from Nvidia, AMD, or Intel. (It's hard to say.) Regardless, having indigenous parts matters more than power efficiency for China right now, and into its future, and we said as much last summer when contemplating China's long road to IT independence. Imagine what China can do with a shrink to 7 nanometer processes when SMIC delivers them — apparently not even using extreme ultraviolet (EUV) light — many years hence....
The bottom line is that the National Research Center of Parallel Computer Engineering and Technology (known as NRCPC), working with SMIC, has had an exascale machine in the field for a year already. (There are two, in fact.) Can the United States say that right now? No it can't.
needless competition (Score:2)
Can the United States say that right now? No it can't.
ok, good job. I hope you do good science with it.
Re: (Score:2)
December 2010:
US Air Force connects 1,760 PlayStation 3's to build supercomputer
https://phys.org/news/2010-12-... [phys.org]
Base 2 number (Score:4, Funny)
We like base 2 numbers and think that the OceanLight system probably scales to 160 cabinets, which would be 163,840 nodes and just under 2.3 exaflops of peak FP64 and FP32 performance. If it is only 120 cabinets (also a base 2 number),
You keep using that word. I do not think it means what you think it means.
Re: Base 2 number (Score:2)
Exactly, they mean "even" number, but it does not sound so much "computers".
Re: (Score:2)
Big brother says: Every number is a base 2 number, but some numbers are more base 2 than others.
Re: Base 2 number (Score:1)
All your base are belong to us?
Re: (Score:2)
Re: (Score:3)
I wonder if this is caused by a translation error, from Chinese? Not that the article was written originally in Chinese, but the source that they maybe copy-pasted from was.
Re: (Score:3)
I think it's a translation error, all right - but it was a same-language translation happening inside the head of a writer with little technical knowledge who was either interviewing a computer scientist or - more likely - was reading a press release from the computer scientist's organization.
Re: (Score:2)
Computers wouldn't be nearly as useful if some numbers couldn't be base 2 numbers.
Re: (Score:2)
What do they mean? Even numbers? Like 50% of all numbers?
Re: (Score:2)
Re: (Score:1)
> Even numbers? Like 50% of all numbers?
You're thinking of integers, haha.
Re: (Score:2)
Integers, you mean. Negative integers can still be even ("divisible by two without a remainder").
Come on, super computer?? (Score:2)
Have seen "transputer" from some Indian outfit 40 years ago, humongous "computational power" and only thing it can calculate is some simple CFD problem with very specific ultra simple geometry and load conditions.
Re: Cases (Score:2)
The cases in the article picture look like they were copied from IBM [hearstapps.com].
Re: (Score:2)
where do you think IBM gets their cases ?
Re: (Score:2)
Apparently left over cases from "OceanLight."
Re: (Score:3)
In terms of raw computing power these CPUs were part of the fastest computer in the world for a few years. They are designed for supercomputers and use a custom RISC architecture.
You make do with what you have ... (Score:2)
it doesn't take the most advanced chip manufacturing processes to create an exascale-class supercomputer,
"I'm endeavoring, Ma'am, to construct a mnemonic memory circuit using stone-knives and bear-skins."
-- Spock, The City on the Edge of Forever [wikipedia.org]
Old 14nm Tech? (Score:2)
And what does a reticle set cost for this 'old tech'? $1M?
My computer is bigger than yours. (Score:2)
Perhaps China has decided to quietly leak that they are first to true exascale without having to publish benchmark results that might show a slightly better performance figure for a US- based machine. Just something to think about.
Seems to be a rather important point in this whole E-penis comparison.
Found the Mao easter egg (Score:2)
As shown in Figure 9, given an image of a running dog and the text "This is a", the model is encouraged to predict the next word, "cat".
WRONG! The predicted text word should be "lackey". Also, the supplied image is of a cat.
Re: (Score:2)
Actually, the next word should be "running", as in "This is a running dog lackey."
Imagine a beowulf cluster of these ! (Score:2)
Nuff said !
Of all of the changes the future has brought... (Score:2)
...perhaps one of the saddest is that no one seems to have posted here to ask us to imagine a Beowulf Cluster of these. ;)