India should focus on developing lightweight models and distributed AI deployment: IESA President

By Pune Media On Mar 4, 2025

Ashok Chandak, the President of the Indian Electronics and Semiconductor Association (IESA), says the country should work on lightweight LLM models.

Instead of chasing GPU power, the emphasis should be on algorithm development, lightweight AI models, and distributed AI architectures like edge AI. This approach leverages India’s talent pool and addresses the limitations of computing power, turning scarcity into an opportunity for innovation, he says in a virtual interview with businessline. Excerpts:

How is the semiconductor space doing globally — the issues and challenges, particularly in the context of geopolitics and supply chain?

Last year was a bit softer for the global electronics industry. However, this year, 2024, has seen growth, crossing about $600 billion in overall revenue worldwide. The growth is mainly in data centres, GPU-related areas, and memory. Other sectors have been relatively stable or stagnant. Chip demand for AI and memory pricing have contributed to the global semiconductor industry growth.

Inventory correction has occurred in various product categories, with most sectors showing stable run rates. Previous years’ inventory in channels and with suppliers limited overall revenue and production increases for non-memory-related products. This correction is expected to improve in the second half of this year, with various sectors regaining run rates, including automotive, consumer, and industrial IoT.

Despite smaller correction cycles, the long-term trend remains positive, with expectations of the industry going beyond $1 trillion by 2030. The industry reached about $625 billion in sales in 2024, almost 20% more than the previous year, mainly due to AI-related chip sales and memory.

Why do you think demand is slack for gadgets, laptops, phones, and all that?

There has been some inventory correction. Most sectors have not shown major growth in run rates but have remained stable. Inventory remained in the channels and with the suppliers.

What kind of growth is the industry looking at this year?

The industry is looking at around 10 to 12% growth worldwide. Other sectors need to contribute, and memory prices are stabilising.

LLMs and applications around them have a shortage of GPUs and computing power, and the cost of GPUs is a hurdle in developing foundational models and other applications. How do you think the world can tide over this crisis?

The shortage is because one particular supplier is getting most of the business. Everyone is seeking the highest computing power from Nvidia, creating a long queue. Nvidia has also improved its supply chain by using vendors like TSMC and Korean companies. Other players like AMD and Broadcom are beginning to roll out their GPUs and chips for AI computing. While big data centres need the highest computing power, not every application does; even lower computing power is sufficient for robots, medtech, and industrial applications to set up AI engines and run LLMs. These issues should stabilise by the second half of the year, and the next year should improve further.

The US government has export restrictions in place, determining who can get what quantity and how they use it. If the US continues export controls, other countries will need to look for alternatives. Regional cooperation and consortiums among multiple countries are needed. Even if countries develop homegrown GPUs, it will take years to master the computing architecture required for the highest performance.

A GPU programme can be rolled out. It’s a long-term solution to have some kind of deterrent. Matching Nvidia’s computing power will be very difficult, even for American, Korean, and Japanese companies. India should be realistic and focus on doing something, as many applications can still run with limited computing power. Multiple ways exist to run AI with limited computing power without needing a supercomputer.

Where does India stand now in terms of the semiconductor industry?

India is making progress, though manufacturing is still in development. The India Semiconductor Manufacturing Programme projects are going to take time to become operational. India’s demand for semiconductors is expected to reach about $103 billion by 2030. In 2024, the demand was close to $52 billion. The bulk of the requirement will be processors and memory, almost half of the demand.

What are the key issues impacting the semiconductor ecosystem?

India has been behind in manufacturing semiconductors. The ecosystem is not fully ready on the manufacturing side, but chip design and verification are quite good. The design services and start-up community are progressing. The government’s DLI policy and design-linked incentive policy are instrumental in further developing start-ups and small and medium enterprises focused on chip design.

The main focus is on semiconductor manufacturing. The Indian government has approved 5 projects, with more approved by state governments, which will come online in the next couple of years. Most of these are OSATs (outsourced semiconductor assembly and test), along with a fab by Tata and a small fab for power semiconductors. The main challenge is the lack of process technology for manufacturing, making India dependent on global players for investment, technology transfer, and know-how.

The biggest bottleneck is that India does not have the manufacturing processes and know-how and depends on global partners. The government is supportive with schemes and incentives. A new policy may be coming soon, following the current Semicon India Programme 1.0. Another challenge is workforce development, as there is a lack of trained workforce for manufacturing due to the absence of manufacturing processes technology. Training in and outside India is needed to ramp up manufacturing.

How can India manage the chip side of AI, given the major bottleneck?

India should focus more on algorithm development and LLMs. It should not worry too much about GPU power, as it has never been a strength. Developing a special project on GPUs is required. Collaboration and regional cooperation are important, as several countries face export control restrictions. India could take a lead role in developing cooperation or a consortium.

With its talent pool, India should focus on developing lightweight models and distributed AI deployment. Instead of relying on centralised data centres, the concept of distributed AI architecture, such as edge AI, can reduce dependency on high-performance chips. Innovation can happen when there is scarcity and restricted availability of computing power, driving efforts to do more with less. The government supports AI, and India has signed a joint statement on AI.

Published on March 3, 2025

_{Images are for reference only.Images and contents gathered automatic from google or 3rd party sources.All rights on the images and contents are with their legal original owners.}

_{Aggregated From –}

_{Source link}