India Plans To Build Fastest Supercomputer By 2017

First time accepted submitter darkstar019 writes "India is planning to build a computer that is going to be at least 61 times faster than the current fastest super computer, IBM Sequoia. Right now the most powerful supercomputer in India is 58th in the list of top 100 supercomputers. From the article: 'Telecom and IT Minister Kapil Sibal is understood to have written to Prime Minister Manmohan Singh sharing the roadmap to develop "petaflop and exaflop range of supercomputers" at an estimated cost of Rs 4,700 crore over 5 years.'"

Comments

Crore = 10^7

[Anonymous Coward]

http://en.wikipedia.org/wiki/Crore

Nice to see the editors making sensibly proof-read, accessibly written summaries, rather than the usual treasure hunt for the true meaning.

Re:Nonsense

[fustakrakich]

Heh, did you? This what you're looking for?

Considering Moore's law

[SpaghettiPattern]

Considering Moore's law that's just about to be expected.

In 5 years we have 3 x 18 month period. The level of improvement in hardware is multiplies by 2^3. Then I'd expect level of parallelism to affect the process by the same magnitude bringing the total to 2^6 = 64.

PARAM and beyond

[unixisc]

India's current supercomputer - one that it's developed since the 80s - is the PARAM, which has had 6 generations to date. The first was based on the Inmos Transputer, the second on an Intel i860, the third on a SuperSPARC II (and it even had an Alpha variant), the fourth on an UltraSPARC II, the fifth on an IBM POWER 4, and the most recent - unveiled in 2008 - was based on the Intel Xeon (Tigerton 73xx). They are currently working on one that's supposed to break the 1 petaflop barrier (that would be 10 crore crore flops for Indians). So this new announcement would be the successor to that.

So it's not like they're new @ this, and what is impressive is that they've used a wide variety of CPUs from different vendors. For this next one, they might want to do that w/ an Itanium III or a POWER7 (unless POWER8 is anywhere close). It would seem that for that, they might get some Intel/HP expertise to help w/ that. I have no idea how good they are @ writing compilers. But yeah, planning a supercomputer based on this CPU and tossing in enough of them should enable them to achieve that goal. Put Debian on it, and then use it for whatever they need - weather forecasting, nuclear simulations or whatever they want to use it for. A lot of the 52 PARAMs that they've manufactured & sold have been sold to other countries.

I just wish that aside from the Indian government, there were a few companies in India that made supercomputers.

Re:PARAM and beyond

[unixisc]

As the history part of the wiki link shows, India originally wanted to get Crays from the US, but couldn't due to a technology embargo by the US. Which actually was justified @ the time, given that India was one of the main technology partners of the Soviet Union and would voluntarily give them things that the KGB couldn't steal from the West (India originally wanted to buy this in the mid 80s). So it's not like they wanted it for dick waving rights - I doubt that any of the first 6 PARAMs were #1 @ any point in time. Weather forecasting was initially the main thing, followed by other things. The US was @ one time concerned about India's use of nuclear technology for making nukes, but in the last agreement b/w the 2, India agreed to separate out its military usage of nuclear technology from its commercial usage. In any case, since these are home grown supercomputers, India could well use a few for nuclear simulations, and there's nothing the US could do.

As for whether India should have nukes when it has the problem of poverty and related issues, India first lost a war against China in 1962 (started by the latter), and IIRC, China was either a nuclear power then, or became one since. India's initial interest in becoming a nuclear power happened as a result of that. Pakistan, in turn, decided to be one so that it could seize back Kashmir, and today is one, but they have been becoming increasingly extreme over the years since 9/11, and are a genuine threat to India. So India would be criminally negligient to its own people if it didn't put a few supercomputers out there to do the nuclear weapons simulations that they need to to ensure that it has a good deterrent. Of course, w/ an Islamic country, even that may not be a deterrent, but since not all Muslims have suicidal fantasies, having that has some likelihood that they'll never be used.

Re:For these 'fastest' metrics:

[cwebster]

Modeling.

Weather modeling (solving navier-stokes and a few other equations on a discrete cartesian grid or on a spherical grid in spectral space). Add in land surface models, ocean models, data assimilation, chemical processes, and then crank the resolution way up and you need a lot of power.

DNS (direct numerical simulation) -- if you want to simulate a fluid flow with turbulence and you want to resolve the turbulence explicitly you need to have a grid spacing in your model that is smaller than the kolmogorov scale. For some flows this may produce a grid spacing measured in millimeters. If you want any decent sized model domain, this produces a lot of grid points.

Monte-carlo type simulations -- i.e., run a simple simulation but do it 1e50 times to amass a statistical representation of the process.

and lots of other types of modeling. Basically if you have a set of partial differential equations that tell us something and you need to solve them numerically (no analytic solutions, etc) and need to do it on very large domains at high resolution and your neighbor grid dependencies are such that your problem is parallel, then a supercomputer is for you.