You can call it an app; one that can speed up scientific computing for almost any industry. But it’s an app that took four years, tens of scientists, and cracking some serious partial differential equations to build. That’s how Santosh Ansumali likes to describe his complex algorithms. Poker faced, which belies mild paranoia — very few outside Ansumali’s field understand his work — he is willing to simplify it to an extent that “people at least don’t get this wrong”.
His paranoia isn’t misplaced. In a country where programmers are a dime a dozen, India cannot claim even one software product in scientific computing which has either gained critical acclaim or popular use. There have been individual research contributions, but never any product. Well, not until now.
So, what are Bengaluru-based SankhyaSutra Labs and Ansumali trying to build? They’ve written a set of numerical algorithms that make computing faster 30X and, at the same time, reduce hardware requirement by one-tenth.
Let’s break it down some more. Say, you need 100 computers to do a job on a supercomputing cluster, this app can do it with 10-odd computers. And because these algorithms compute so fast, their simulations are far more accurate than what standard tools can achieve. One of the Indian defense forces got a feel of this recently. So have others in the semiconductor, wind energy, automobile, space, and petrochemicals sectors.
A boutique startup’s boutique offering need not follow a boutique flight path though.
At the beginning
In the late ‘90s, Ansumali was a master’s student at the Indian Institute of Science while his friend Ronojoy Adhikari was pursuing his PhD. Their respective supervisors received one of the largest projects in fluid dynamics from Unilever. Calculating the flow of fluids has been one of the toughest supercomputing problems. The two students got a chance to work on the Lattice Boltzmann equation — a technique those days to solve fluid dynamics problems. They wrote some code, solved some problems for Unilever and went off to pursue their doctorate and post-doctoral research in Switzerland and Edinburgh respectively.
They returned to India in late 2000s. As they settled into academic positions they continued to explore problems of real world simulation, thinking it would be a radical change if they could remove some approximations used by engineers. By then Ansumali had met Sunil Sherlerkar, then a director at Intel Labs. Computational fluid dynamics is a big consumer of high-performance computing.
Around 2011, the chip giant was looking for software that could extract the real speed from supercomputers. Even in the best of cases, high speed computers perform at 20-30% of the stated speed, or floating point operations per second. What happens is, when you start scaling software using more and more processors, the speed increases at first but after sometime it flattens or even drops. That’s because the communication cost for exchanging data between processors starts becoming the bottleneck.
