Global AI Conclave: India needs to build frugal LLMs, has to opt for ‘heterogeneous’ architecture, says Intel’s Santhosh Viswanathan

Global chip maker Intel’s India MD Santhosh Viswanathan said that India needs to build its own foundational large language models (LLMs) but frugally to leap frog into the ongoing AI revolution.

His views come at a time when India’s biggest tech voices including Nandan Nilekani and Narayana Murthy have been talking about the country becoming AI use case capital of the world, leveraging the foundational models of global tech giants Google, Open AI and Microsoft instead, as they are expensive to build.

Speaking at CNBC-TV18 & Moneycontrol Global AI Conclave 2024, Viswanathan said, “We definitely need Indic languages, local cultural context (in LLMs). But majority of the LLM models that are used, come from outside of India. We have work to do to get there. And we have to start to think how do I make LLMs in the most cost-effective manner and not follow the whole path of high capital investment and infrastructure.”

Highlighting the cost structures and high capital required, he added that many of the leading tech companies have spent about $100 million plus to go train a frontier model.

“That's the minimum amount of expenses that is required. It’s not just financial cost but also the environmental cost. One megawatt of data center requires about 69,000 liters of water. And if you take 200 megawatts (of data centre) that probably Bangalore has, that can basically go and give water to about 1.5 lakh homes in Bangalore,” he added.

Viswanathan suggested that India will need to focus on a ‘heterogeneous’ architecture to develop LLMs frugally, which will involve a combination of CPUs, GPUs and LLMs and small language models (SLMs) depending on the use case required. He also added that India needs to play on its unique strengths of resources as it has 16 percent of the world’s AI talent and population scale innovation.

“We have taken multi-year big bets in semiconductor manufacturing, AI Mission to create capacity that we very much need. Whether it is state or Central government, all of them have the right intention that this is the leap frog moment for India. But one of the things I really worry about is can this actually be an India moment, because we follow the path that the west lays out,” Viswanathan said.

Citing an example about large data centres, he said that India produces 20% of the world’s data, and are 3 % of the world’s servers so the country is in that pursuit of building last data centres. Though the country still needs large data centres but even tripling the capacity gets us to about only 4% of where the US is today. In fact, Northern Virgnia has 2X capacity of all of India today, Viswanathan shared.

CNBC-TV18 & Moneycontrol Present Global AI Conclave 2024, Co-Powered By Intel AI, Associate Partners Adobe & Reliance Industries Limited, Ecosystem Partner TiE Bangalore, and Event Technology Partner Townhall