70 supercomputers by 2022: Great Vision, But Questions Over Execution

The Indian government’s plan to build 70 supercomputers each capable of delivering up to 50 petaflops of performance sounds ambitious indeed. Considering that the fastest supercomputer India has today is a 0.79 petaflop IBM machine, this will be a quantum leap if those in charge are able to bring the project to fruition.  

The project is looking at building machines faster than the fastest functional supercomputer—Tianhe-2, the Chinese supercomputer that delivers almost 34 petaflops.   

There are of course faster supercomputers being built such as the 50,000 node,180 petaflop Aurora which Intel and Cray are building for Argonne National Laboratory in the United States. But these are standalones and will take a few years to build.  In contrast, the Indian project is looking at a combined floating-point performance of 3500 petaflops.

While the finer details of this project haven’t emerged, we know that the government is roping in institutions like the Indian Institute of Science (IISc) and Center for Development of Advanced Computing (C-DAC) for the project.

It’s not clear at this point in time whether these institutions will actually build the systems or they will evaluate available technologies and commission the project to one are more players in the High Performance Computing (HPC) space. If past experience is any indicator, getting Indian research institutions to build the machines within a deadline may not be a great idea.

 In the late eighties, when the United States restricted supercomputer sales to countries like India, the Indian government funded research institutions like National Aeronautical Laboratory (NAL) and Center for Development of Telematics (C-DoT) to develop parallel processing machines using commercially available processors. NAL and C-DoT both built special purpose parallel processors. Moving forward, the government established C-DAC to build a general-purpose parallel processor. The institute was sanctioned Rs.36 crore and given a deadline of three years to put together a 64 megaflop machine.

Param, the machine delivered by C-DAC three years later technically met the required specifications, but turned out to be a dud when it came to running applications. In the following years, the institute came up with later versions of Param with more powerful processors, but by then user organizations had other options such as IBM as the US government had relaxed curbs on HPC exports to a certain extent.

The other scenario that the Indian government will have to consider before embarking on the project is the current US stand on exporting HPC equipment. Recently the US restricted Intel from supplying processors to China that were to be used in the development of a faster supercomputer. As of today all seems to be well between India and the US in the new geopolitical order and a slew of MoUs have been signed between the two countries for cooperation in science and technology.  Over the last few years the US has allowed its companies to sell HPC solutions to various institutions. But here we are talking of an insanely high combined floating point performance. Considering that India is planning to use the computing power for defense applications will the US allow India access to the technology? 

The vision of building the supercomputers sounds great. It’ll be interesting to see how the people in charge of the project will address these issues.