Slashdot Mirror
China Makes World's Fastest Supercomputer
shmG writes "China has replaced the United States as the maker of the world's fastest supercomputer. A Chinese research center has made the world's faster super computer — named as Tianhe-1A, which was released at a national conference on high-performance computers (HPC) in China. Made at a cost of over $88 million, Tianhe-1A is theoretically able to do more than 1 quadrillion calculations per second (one petaflop) at peak speed. Tianhe-1A 's peak performance reaches 1.206 petaflops, and it runs at 563.1 teraflops (1,000 teraflops is equal to one petaflop) on the Linpack benchmark."
Stolen? I don't know. Purchased? From the article:
So unless Nvidia and Intel have reported 20,000 or so stolen processors lately, I wouldn't worry too much.
W..w..W - Willy Waterloo washes Warren Wiggins who is washing Waldo Woo.
Serious research still needs much faster supercomputers than we have now. All kinds of science - from artificial intelligence to weather modeling to astrophysics to genetic research to nuclear simulations. Access to a powerful supercomputer is a major boon for academia in the country.
As someone whose works depends on HPC I disagree with you. A lot of people in life sciences, materials science, nuclear physics, geophysics and other knowledge areas needs clusters and super computers.
English is not my first language. Corrections and suggestions are welcome.
Oak Ridge (Jaguar):
Cores Rmax(GFlops) Rpeak(GFlops) Nmax Nhalf
224162 1759000 2331000 5474272 0
Seems faster by a good margin.
The slashdot summary has the wrong numbers. The actual article which slashdot quotes is contradictory. Its starts by saying:
"Tianhe-1A has set a new performance record of 2.507 petaflops, as measured by the Linpack benchmark, making it the fastest system in China and in the world today."
and then one paragraph later it gives the same numbers as the slashdot summary.
Other articles (from other sites) are claiming theoretical peak performance of 4 Petaflops (from an Nvidia source) and sustained petaflops of 2.5.
The article starts by claiming 2.507 petaflops, but gives no mention if that is Rmax or Rpeak. We have to assume that it is Rmax, since 2.5 petaflops is no big deal in terms of Rpeak.
Unfortunately, then the article lists both Rpeak and Rmax. But the numbers quoted seem to be for Tianhe-I (#7 on the top 500 list), not Tianhe-IA (not currently listed). Wikipedia table of the top 10
Oh, and it gets better. The article claims that Tianhe-IA has 7,168 GPUs and 14,336 CPUs. Very strange, since the Tianhe-I has 71,680 CPU/GPU pairs.
My guess is that China doubled up their Tianhe-I computer and swapped out for newer GPUs, then named the new thing Tianhe-IA (this is pretty normal when competing for top500 spots). I'm going to go with 143,360 Xeon/M2050 pairs. Either that, or the Chinese found a way to overclock 10% of their chips into the 20+ GHz range and threw out the rest.
See that "Preview" button?
Nearly all of these supercomputers are just that - VERY large clusters.
Although in many cases they have specialized communications backplanes for communications between nodes with capabilities (such as low latency) that can't be achieved with geographically distributed clusters. (Note the mention of parts from Intel and Nvidia, combined with undefined "domestic" communications silicon.)
Also note that geographic distribution leads to all sorts of information assurance nightmares when you're simulating nukes...
retrorocket.o not found, launch anyway?
Is this faster than Blue Waters at NCSA is going to be in 2011?
Nope, Blue Waters is supposed to be significantly faster. According to NCSA's page about Blue Waters, Blue Waters is supposed to have peak performance of 10 petaflops, and sustained performance at 1 petaflop. Tianhe-1A, according to the summary, peaks at about 1.2 petaflops.
IBM and the NSF are already building a supercomputer which will have an eventual supposed peak performance of 10 petaflops: Blue Waters http://en.wikipedia.org/wiki/Blue_Waters At the university of Illinois. I actually work at the university and have been by the building (it is complete), though we don't have the hardware yet from IBM. It is all starting to go in, though, and is supposed to be working before October of next year. So don't worry, it's not like the US is getting stomped in computing power.
Also, to those posters questioning the need for supercomputers......as someone who works on HPC code and applications, I cannot disagree with you more. We already have simulations that are not practical to run on modern supercomputers. In the grand scheme of things, these are not even terribly "complex" issues when compared to, say, modeling an entire human organ or even body. Imagine what we could do with something like that....you could test new "drugs" in a simulated environment and get a good idea of the effects (this is all just one idea that comes to mind)...
Yes, this is spot on for massively parallel systems. The interesting thing is that China does actually make their own interconnects, but they aren't so great. The Tianhe-1 actually runs at 47% of the theoretical capacity. In contrast, the previous number 1 (Jaguar) runs at about 76%. In fact, China's previous big HPC was Nebulae, which had a higher theoretical peak than Jaguar, but didn't actually perform faster because of interconnects problem.
Gentlemen! You can't fight in here, this is the war room!
I've not really invented anything... but I have re-imagined what FPGAs could be, if someone were to toss out routing and go as fine grained as feasible.
Been done too. One of the commercial FPGA architectures (the name escapes me at the moment) in the early '90s was essentially what you describe. It didn't prove to be particularly effective or useful and the company went out of business.
I'll have to dig around and see what the name was.