New Delhi, Sept. 26: India’s plan to build what could be the world’s fastest supercomputer by 2017 is unrealistic and at least three years out of step with a timeline pencilled by Indian and foreign computer scientists.

Scientists say the plan for an exascale supercomputer capable of a billion billion floating point operations per second (flops), a performance measure for computers, ignores key technology hurdles that make dreams of such a machine before 2020 unrealistic.

The plan for a homegrown exascale machine to tackle complex unsolved problems in weather forecasting, drug design, and engineering research, among other applications emerged from the Union information technology ministry last week.

IT minister Kapil Sibal is reported to have written to Prime Minister Manmohan Singh outlining a Rs 4700 crore proposal from a government supercomputer laboratory to develop petascale and exascale machines over the next five years.

But scientists caution that even if such a behemoth machine is built with present-day technology, it would need more power than what is supplied by India’s largest nuclear reactors, generating 540MW of power.

“A modern supercomputer usually consumes between 4 and 6 megawatts — enough electricity to supply something like 5000 homes (in the US),” Peter Kogge, a computer scientist at the University of Notre Dame in the US wrote in the journal IEEE Spectrum last year.

Kogge, the leader of a panel of experts asked by a US defence research agency to examine technology hurdles towards creating exascale machines, has described power requirements as the biggest obstacle. An exascale supercomputer would need 1500MW power, he wrote in the journal.

The panel’s report also lists the lack of available technology to support memory and storage operations on exascale machines, among three additional “major challenges.”

Scientists at India’s Centre for Development of Advanced Computing, (CDAC) the very laboratory that has proposed the exascale machine said they are “aware’ of these hurdles and plan to spent the next five years only on research to understand them.

“We’d like to initiate research and development now so that by 2017 we understand what (technologies) an exascale supercomputer would require and we can start building one,” CDAC’s executive director Rajat Moona told The Telegraph.

The scientists said the 2012-17 period will be used to build petascale computers — a petascale machine performs a million billion operations per second, 1000 times slower than an exascale system.

The world’s top 10 high performance supercomputers are located in the US, Japan, China, France, and Germany. The proposed exascale effort, Moona said, is intended to help India develop a homegrown exascale machine on par with international efforts by 2020.

The fastest supercomputer in the world today is at the Lawrence Livermore National Laboratory in the US, an IBM machine named Sequoia with a speed of about 16 petaflops.

In contrast, India’s fastest machine at the Centre for Mathematical Modeling and Simulation, Bangalore, crunches data at 303 teraflops (a thousand billion flops), and the CDAC’s fastest is named Param Yuva rated at 54 teraflops.

The proposal for the Rupees 4700 crore effort is a “draft conceptual proposal” that outlines only research efforts on exascale technologies over the next five years, a source in the information technology ministry told The Telegraph.

One major task ahead is to work on tailoring applications so that computational tasks can be broken up and distributed across the parallel processors that make up exaflop machines, a computer science engineer at CDAC said.

“We can ask 100 people to dig a large pond in two days, but we can't put 100 people to dig a well,' said Pradeep Sinha, head of CDAC's high performance computing division. 'Some applications will need to be redesigned from scratch,' Sinha said.

Another challenge, Sinha said, is building such a giant system that is resilient, or fault-tolerant. “We may have to examine novel self-healing technologies, he said.

Supercomputers are already routinely used in weather forecasting, genome studies, aerospace simulations and drug design, among other applications. Faster supercomputers, Sinha said, will help increase the accuracy and reliability of such applications.