India's AI vision calls for 80 exaFLOPS of infrastructure

The India AI group at the nation's Ministry of Electronics and Information Technology (MeitY) last Friday published an AI vision document that calls for a massive build of national computing infrastructure.

The planned infrastructure will comprise 80 exaFLOPS of power across three layers – high-end compute, an inference arm, and edge compute – enabled by a distributed data grid with speeds of 200/400 Gbit/sec.

The high-end compute, a resource intended primarily for training models, is suggested as packing 40 exaFLOPS of power drawn in part from 10,000 GPUs – plus 200 PB storage. The vision document calls for the mid-range tier, or inferencing arm, to have 20.8 exaFLOPS (12 for training and 8.8 for inferencing) spread across four geographical centers in India, plus 400 PB storage. Each of the four centers needs 1,000 GPUs for inferencing tasks and 750 for AI training.

"For users with limited compute resources and smaller datasets, the edge compute resource proves invaluable. It allows users to test their hypotheses on low-end machines and their own data initially," states [PDF] the document. The edge centers receive 240 PB of storage and are spread out in 12 different places across India. Each center has 125 GPUS of AI training totaling six exaFLOPS, and 500 GPUs for AI inferencing totaling 13.2 exaFLOPS. Its unclear if those exaFLOPS are expected on each site, or across the edge fleet.

• China updates national computing plan with calls for more edge, storage, memory, and … Blu-ray?
• The future of the cloud sure looks like it'll be paved in even more custom silicon
• India demands social networks 'swiftly' remove all CSAM
• Big Tech offers free training courses on India's new digital skills platform

India's planned investment in 80 exaFLOPS is enormous, but also rather less than the 150 exaFLOPS of additional power that China announced earlier this month. That effort is backed by national storage capacity of 1,800 exabytes. Details on Beijing's plans otherwise have been vague.

India AI did note that China has the largest share of the world supercomputers – at 162 as of November 2022, comprising 32 percent of machines whose details are recorded on the Top 500 list. India, meanwhile, only counts three – but the nation's National Supercomputing Mission (NSM) aims to increase this number to 24.

The document also calls for assembly of national data training resources, including non-personal data, under a program known as the India Datasets Platform (IDP).

The datasets have potential for data-driven governance, as well as for startups and research, said India AI.

A working group designated to further plan IDP would also be in charge of proposing a pricing model to access the data – with fixed, one-time, and subscription options available.

The AI vision document is just that, so there's no certainty it will be followed, despite ministerial enthusiasm for India to adopt AI at speed. It's also silent on procurement – a significant issue given that sourcing GPUs is not easy for any entity. ®