India's Leap in Supercomputing: Innovating for Tomorrow

As India strides towards self-sufficiency in supercomputing, embracing this evolution isn't just an option-it is pivotal for global competitiveness and technological leadership.

- Pankaj Panjwani

Supercomputing architecture is undergoing a transformative shift globally, driven by the need for unprecedented computational power. High performance compute (HPC) machines stand out for their ability to perform complex calculations at speeds far beyond conventional computers. Nations race to harness these capabilities for advancements in weather forecasting, drug discovery, artificial intelligence, and more. With significant strides in indigenous motherboard design and local manufacturing, coupled with ambitious initiatives like the National Supercomputer Mission, India’s journey in supercomputing is the one to watch.

High performance compute (HPC) vs general compute

By definition, an HPC machine must be able to do 10x more computations than the fastest laptop or desktop available at that time.

An HPC machine must process multiple operations in parallel to achieve high computational capacity. A general-purpose computer, like a laptop, has much lesser parallelism compared to an HPC machine. The silicon chip used in general-purpose machine is designed with very strict thermal constraints (that a person can withstand) and power dissipation constraints (with which the battery must last longer).

Whereas for an HPC, although the thermal and power constraints are targeted, the biggest objective is getting maximum performance. The performance capability is measured in terms of FLOPS (floating point operations per second).

A silicon designed for high performance compute must be able to perform multiple operations in parallel to achieve higher number of compute operations in a second. The hardware architecture and compiler play a critical role in achieving such parallel processing.

The applications for HPC are much different than a general-purpose compute. An HPC application will typically have workloads as a single operation/calculation on multiple datapoints.