Slashdot Mirror
US Once Again Boasts the World's Fastest Supercomputer (zdnet.com)
The US Department of Energy on Friday unveiled Summit, a supercomputer capable of performing 200 quadrillion calculations per second, or 200 petaflops. Its performance should put it at the top of the list of the world's fastest supercomputers, which is currently dominated by China. From a report (thanks to reader cb_abq for the tip): Summit, housed at the Oak Ridge National Laboratory (ORNL), was built for AI. IBM designed a new heterogeneous architecture for Summit, which combines IBM POWER9 CPUs with Nvidia GPUs. It has approximately 4,600 nodes, with six Nvidia Volta Tensor Core GPUs per node -- that's more than 27,000. The last US supercomputer to top the list of the world's fastest was Titan, in 2012. ORNL, which houses Titan as well, says Summit will deliver more than five times the computational performance of Titan's 18,688 nodes.
Have gnu, will travel.
If you had access to a computer this powerful, what would you do with it? just asking...
"Knowing everything doesn't help..."
If you had access to a computer this powerful, what would you do with it? just asking...
. . . in other news . . . the DoE has just announced that they own 51% of the Bitcoin Universe . . .
Schroedinger's Brexit: The UK is both in and out of the EU at the same time!
This is kinda silly. The fastest supercomputer is going to be whomever has built the newest one. Moore's law is slowing down a tiny bit, but it's still going. Someone spending $100,000,000 on a supercomputer today is going to have a slower machine than someone building one for the same amount in a year's time.
While I agree it is silly, the top 10 supercomputers are not in chronological order. Their commission dates and ranking are listed below. The top 5 newest supercomputers in this list are #1, 3, 4, 8, 9 in PFlops. Considering the 5 newest supercomputers average a rank of #5 and the 5 oldest supercomputers on the top 10 list have an average rank of #6, there is little correlation between age and computational power among the top machines. That has far more to do with what the machines were built to do and what budget they had, not the year they were built. Being newer does make your dollar stretch much further though.
1) 2016
2) 2013
3) 2016
4) 2017
5) 2012
6) 2013
7) 2015
8) 2016
9) 2016
10) 2011
-- All that is necessary for the triumph of evil is that good men do nothing. -- Edmund Burke
There are several very useful for humanity topics, requiring lots of calculations: ;)
- fusion, whether for tokamacs or stellarators
- proteins for medicine
- properties of alloys
- PI (to see if Carl Sagan was joking or was he on to something
The fastest supercomputer is going to be whomever has built the newest one.
Not necessarily. IBM is also building a newer . . . slower . . . supercomputer for the DoE:
Summit is one of two of these next-generation supercomputers that IBM is building for the DEO. The second one is Sierra, which will be housed at the Lawrence Livermore National Laboratory. Sierra, which is also scheduled to go online this year, is less powerful at an expected 125 petaflops, but both systems are significantly more powerful than any other machine in the DoE’s arsenal right now.
Schroedinger's Brexit: The UK is both in and out of the EU at the same time!
Nobody is going to win in Somalia, that's for sure, but it still takes tech-know how, not just $$$ and timing. HPC is NOT like building a gaming PC, man.
I'd rent it out. Sell processing time on it. How else would one pay the power bills? You can rent a lot of processing power on Amazon, but you can't rent a supercomputer-grade high performance interconnect between your nodes. Not from them, anyway.
I'd offer a special service for non-profit customers: They can pay power costs only, but then they only get to use idle resources. Commercial-rate customers take priority - got to pay the bills somehow.
How is this any different than a data center of the same scale? I'm not even sure it has a higher bisection bandwidth.
Okay—I am sure it will have way lower latency between nodes at any equivalent bandwidth tier. But unless you're planning to aggressively exploit that, it's just your garden variety datacenter make more expensive and less flexible.
1977 Heinz Ketchup Commercial "Anticipation"
Heck of a lot of money, Brownie, for less ketchup.
How about making a Beowulf cluster of those? Oh, what a minute. ;-/
LoB
"Anyone who stands out in the middle of a road looks like roadkill to me." --Linus
Someone spending $100,000,000 on a ANYTHING today is going to have a slower machine than someone building one for the same amount in a year's time.
Fixed that for you. You could say the same about anything really, because everything is getting faster, smaller, lighter, etc.
my karma will be here long after I'm gone
While I agree it is silly, the top 10 supercomputers are not in chronological order. Their commission dates and ranking are listed below. The top 5 newest supercomputers in this list are #1, 3, 4, 8, 9 in PFlops.
But he said if they spent the same amount next year it would be faster, which is true. We don't know the prices of the machines you listed.
my karma will be here long after I'm gone
My Intel desktop has the performance of a 90's supercomputer. Strange to think the best technology available to the world's top scientists in the 90’s now sits on my desk
my karma will be here long after I'm gone
The supercomputers are measured using Linpack. This is a simple benchmark and can be done efficiently by large number of dumb cores in parallel. Though some workloads mimic Linpack, most workloads don't, so Linpack is not a good measure of supercomputer speed. Another benchmark which is gaining momentum nowadays is HPCG (see http://www.hpcg-benchmark.org/ ) which measures more broader performance and it shows quite a bit different picture. The top is still the 2011 K Computer. Some NEC computers which don't make in top500 are ranked in thirties. The Chinese computer is current top500 leader but it is ranked 5th on HPCG, even behind their own other computer which is 1/3rd the speed on Linpack.
A giant nethack dungeon with thousands of players playing at the same time fighting each other!
BTW, a fun detail from the article that should really have been included in the summary:
James Hack, director of ORNL's National Center for Computational Sciences
Escher was the first MC and Giger invented the HR department.
Quite a few systems lasted more than six months at number one. among them
ASCI Red
Earth Simulator
BlueGene/L
Numerical Wind Tunnel
Roadrunner
Tianhe-2
Sunway
And some computers were only on top for an instant.
everything is getting faster, smaller, lighter, etc.
My wife's sister is not.
You can rent a lot of processing power on Amazon, but you can't rent a supercomputer-grade high performance interconnect between your nodes. Not from them, anyway.
https://aws.amazon.com/hpc/
Supercomputers these days are superscalar, so you can add as many racks and shelves of nodes as you have space and money for. When they do upgrades, they just swap out the old slowest nodes and add new nodes with the faster CPUs and GPU's.
Vintage computer adverts: http://www.vintageadbrowser.com/computers-and-software-ads
Good players at tic-tac-toe will always end up in a draw.
But you don't even need an electronic computer to achieve this. The Tinkertoy computer was made entirely out of string and balsa wood:
http://www.computerhistory.org...
Vintage computer adverts: http://www.vintageadbrowser.com/computers-and-software-ads
t could have been built by any six-year old with 500 boxes of tinker toys and a PDP-10.
what sort of child has a pdp-10?
Ssshhhiiiiittttt. Don't put it past the Gov to pull a stunt like that.
Step 1. Print money via the Treasury in addition to spending tax payer money.
Step 2. Build supercomputer
Step 3. Mine Crypto Currency (Bitcoin)
Step 4. Sell said Crypto Currency.
Step 5. Recycle the money back into the Gov.
Life is not for the lazy.
The bandwidth in Amazon's HPC nodes is sort of ok compared to some cluster configurations, but the latency still leaves a fair bit to be desired. Last I tested, 6 months ago, the AWS HPC networking still had a 6 times higher latency on real world tasks I tested it with, compared to Infiniband equipped nodes(depending on the node hardware in question, GPU's might even talk over Infiniband to another GPU inside the node, if it's on another PCIe root complex, because it's faster, and I didn't receive an answer to that question from Amazon when I asked)
Rent it to the IRS and DoJ.
Ensure every cyber currency in the USA was interacted with using the big gov super computer.
When the cyber criminals want to cash out their digital currency collection, the US gov is ready to help with that transaction.
Domestic spying is now "Benign Information Gathering"
? You don't seem to be the only one with this misconception. I see several posts below that echo you. Note that this supercomputer temporarily replaces a Chinese supercomputer at the top of the list that did not use American processors. So, prior to this one, the world leading supercomputer did not use American processors.
Furthermore, this one is not an Intel or AMD machine. Still American companies, but each node in this one has 2 IBM Power9s and 6 NVIDIA V100 Voltas. In fact, the GPUs are the real compute power in most of the really big supercomputers these days and lots of them run NVIDIA GPUs.
The Power9s appear to be made in Malta, NY, but the NVIDIA V100s, arguably the real compute power, are made by TSMC in Taiwan. This machine might be designed in America, but it appears to be mostly made in Taiwan.
These processors are made to work together. The Power9 has native support for the NVIDIA's NVLink connections. The Power9's 16GB of HBM2 memory maintains cache coherency with the Voltas, essentially acting as an L3 cache for the Voltas.
The core count in this system is mind-blowing. The Power9s are apparently the version with 24 cores each. So 24*2*4600 = 220,800 SMT4 cores. The Voltas are hard to quantify. Each Volta has 84 Streaming Multiprocessors (SMs). By that count, there are 84*6*4600 = 2,318,400 SMs. This is where the real supercomputing occurs. But each SM in turn contains 64 FP32 cores, 64 INT32 cores, 32 FP64 cores, 8 Tensor Cores, and 4 texture units. I'm not going to even bother doing the math. Suffice it to say that this machine has already broken the exaflop barrier in problems that didn't require FP64 precision.
The transistor count is also astounding. The Voltas have 21.1 billion transistors each. The Power9s have 8 billion. ((21.1*6)+(8*2)) * 4600 = 655.96 trillion transistors just in terms of processors. We've come a long way since the Intel 4004's 2300 transistors less than 50 years ago.
IBM POWER and PowerPC get short shrift.
We can only imagine how many fewer datacenters we'd have if these pure RISC implementations with superior multiprocessing capability were used instead of x86-64 emulating crusty old instruction sets.
Kriston
Azure is competing in this space along with AWS.
I am shocked that these firms don't yet offer to bill by CPU usage. I come from the end of the age when CPU cycles were charged per contract and were planned using budgets. It wasn't for rationing but to help pay for the resources.
Kriston
No, the Power9 is fabbed in New York.
odd.. because your illiteracy rate hasnt changed in 10 years, and is higher then china's with their "primitive chicken scratch"
https://www.huffingtonpost.com...
Care to explain that one for us?
The story has focused so far on how the US got the #1 crown back. But the real story is about how we can now run the fastest and largest AI jobs. Because this IBM supercomputer has 27K+ GPUs, it can run massive deep learning jobs. IBM has been very focused on this deep learning space with their TensorFlow-based open-source PowerAI software offering.
it states it is running Red Hat Enterprise Linux (RHEL) version 7.4. System User Guide - Overview - OS
My beliefs do not require that you agree with them.