India aims to play catchup in supercomputing to support AI in government services

As spending on generative AI continues at an unprecedented rate, the Indian government is looking to adopt the new technology to improve government services and support research initiatives for its citizens.

IndiaAI, a  a body created by India’s IT ministry, has released a report that seeks to not only lay down a framework for data management for use of generative AI in government services, but also increase the country’s supercomputing infrastructure.

As part of the report, the expert committee has suggested that the country needs to set up an initiative, dubbed FutureLab AI, to upgrade its existing compute infrastructure for AI via a public-private partnership model.

“The overarching infrastructure comprises of the three-layered systems, high-end compute infrastructure, inference arm, and edge compute that are strategically distributed to meet users’ computational requirements efficiently,” the committee wrote.

High-end compute, according to the committee, refers to systems that will be used for handling computationally intensive tasks like AI model mixed-precision training, pre-training and fine-tuning, and large-scale model training.

For this purpose, it has recommended the use of what it describes as a tier-1 facility, which would  boast 40 Exaflops of AI computing power, driven by 10,000 graphics processing units (GPUs) in a single cluster.

The tier-1 facility is expected have high-performance storage of 200 petabytes.

For a second-tier facility, which would service inferencing, or operational, processes, the committee has suggested the use of 3 Exaflops of AI computing power, driven by 750 GPUs. This tier is planned to support 400 petabytes of high-performance storage.

The second tier, according to  the committee, should be strategically distributed across four geographical centers in India, in the north, south, east, and west.

“Tier 2 centers support both AI training and inferencing tasks for their respective regions, reducing latency and improving accessibility to AI resources,” the committee wrote.

For edge computing, the committee suggests the use of 500 Petaflops of AI computing power, driven by 750 GPUs for AI training and an additional 500 GPUs for inferencing tasks. It has also recommended that these systems be placed across 12 geographically diverse regions.

Altogether, the committee has recommended the use of 14,500 GPUs with 58 Exaflops for AI training performance and 10,00 GPUs with 22 Exaflops for AI inferencing performance, bringing the total count to 24,500 GPUs with 80 Exaflops of computing power.

The total high-performance storage recommended by the committee stands at 840 petabytes.

India lags in supercomputing power

However, when compared with other countries, such as the US and China, India is far behind in terms of supercomputing infrastructure that can support AI-related workoads, a report from OCED quoted by the expert committee showed.

While India presently has just 3 supercomputers, China and the US have 162 and 127 supercomputers, respectively.

“The People’s Republic of China has the largest share, accounting for 32% of the top supercomputers. Following China, the United States holds 25%, Germany 7%, Japan 6%, France 5%, and the United Kingdom 3% of the top super computers,” the committee wrote.

The committee has also suggested the use of a data management framework, dubbed India Data Platform (IDP).

The IDP, according to the committee, is expected to contribute to the growth of India’s AI ecosystem by providing access to diverse and quality data sets, facilitating data collaboration, enabling data-driven AI models, accelerating AI research and development, and enabling the government to become a data-driven organization.