Takshashila Issue Brief - AI Race Must Not Overlook AI Hardware
Artificial Intelligence (AI) has been one of the most transformative technologies of the 21st century, and it is expected to continue to revolutionise the way we live and work in the years to come. The impact of AI is being felt across industries and sectors, from healthcare and finance to transportation and manufacturing.
The last few months have seen multiple AI applications built on large language models (LLMs) being launched worldwide with Chat GPT cornering most of the attention. Tech giants like Microsoft and Google have committed to adopting AI applications across their service offerings.
The increasing sophistication of AI algorithms and the exponential growth of data has opened up new possibilities for AI applications. Companies and governments are investing heavily in AI research and development to stay ahead in the race to harness the power of this technology.
While much of the popular AI discourse has focused on software applications of AI, it is equally important to not lose sight of the other side of AI technology: AI hardware — the computing infrastructure powering AI and its functions.
In this Issue Brief, we look at our past research on enabling India to become a leader in AI hardware manufacturing.
AI Hardware: Types and Applications
AI hardware includes field programmable gate array (FPGA), graphic processing unit (GPU), and application specific integrated circuit (ASIC).
While AI applications can run on existing and ordinary hardware, semiconductor companies are building next-generation accelerator architectures with increased speed, computational efficiency, and large data set processing abilities. This specialised AI hardware can supercharge the AI race by serving as a differentiator in AI applications.
In fact, AI-related semiconductors are set to grow at 18% over the next few years as opposed to the 2% projected growth for non-AI ones, and by 2025, they can account for 20% of the total demand for semiconductor chips.
The shift in demand from general-purpose to AI-specialised chipsets is creating new opportunities in the semiconductor market. Strategic sectors like defence, space, and telecommunications are likely to employ custom AI-specialised hardware.
In broad terms, AI applications pass through the stages of training and inference - exposure to large data sets for the algorithm to learn, followed by responses to user inputs. Training and inference have varying demands on AI hardware. Hence, AI-enabled hardware can be classified into two types, training and inference chipsets, and this creates distinct opportunities for manufacturing AI-specific chipsets.
AI Hardware Policy for India
India is making a huge push toward semiconductor manufacturing and participation in global chips supply chains. Under the Chips to Startups programme launched by MEITY, India is already looking to build application-specific semiconductor chips
AI inference chips can prove to be a good bet for India in building its semiconductor ecosystem. Inference chipsets are low-cost and do not require the technological, IP, and manufacturing capacity that training chipsets do.
In Takshashila Discussion Document - An AI Hardware Ecosystem in India, we undertake a SWOT analysis and make the following recommendations for an AI hardware policy for India:
A Fab for Inference Chips Production: The Union Government’s semiconductor package of 2021 includes a scheme for financial incentives for building fabs in India. Priority must be given to building a fab for producing AI inference chips, perhaps under the public-private partnership model.
Fund and Support Open Source Projects Related to AI Hardware Design: To circumvent proprietary ownership of technology in this field, the government must fund and support open-source projects to design AI hardware. IISC’s Aryabhat-I is a ready contender.
Include AI Hardware in Existing Policy Schemes: Scheme for setting up Compound Semiconductors Facilities and Design Linked Incentive (DLI) for the semiconductor industry must be expanded to include AI hardware and its design aspects.
International Collaboration with the U.S. and Quad: Multilateralism is essential to gain a strong foothold in semiconductor supply chains. Technical collaboration with the U.S. will help India. The Quad’s Semiconductor Supply Chain Initiative’ must also include AI hardware as part of its collaborative efforts.
This Issue Brief has been compiled by Shrikrishna Upadhyaya, with inputs from Pranay Kotasthane.
Further Material on AI from Takshashila Institution:
Takshashila Discussion SlideDoc - An AI Hardware Ecosystem in India: A SWOT Analysis by Arjun Gargeyas, Samparna Tripathy and Anup Rajput
Takshashila Discussion SlideDoc - Putting India on the road to becoming an AI superpower by Takshashila Bangalore Expert Group
Takshashila Working Paper - China's Quest For AI Leadership: Prospects and Challenges by Manoj Kewalramani
The race for the domination of AI chips - Hindustan Times by Arjun Gargeyas
Rebooting AI in India — The Takshashila Institution by Shailesh Chitnis
India must dominate the game of chips – through its human resources | Mint by Nitin Pai
Takshashila Report - An Open Tech Strategy for India (A Working Draft) by Apar Gupta, Arjun Gargeyas, Bharath Reddy, Kailash Nadh, Nitin Pai, Pranay Kotasthane, Rushubh Mehta, Saurabh Chandra and Venkatesh Hariharan
Takshashila Working Paper - High-Technology Geopolitics in the Post-Pandemic World by Pranay Kotasthane
