The Centre’s decision to subsidise the cost of high-end computing is expected to usher in a wave of research in Artificial Intelligence (AI) and related areas from the funds-starved academic sector. High cost of processing AI workloads and the fact that Big Tech has been draining academia of cream talent has kept academicians on the sidelines of AI research and development. However, this is likely to change now with the government subsidising GPUs, which makes up the largest cost in AI research.
Take for instance, the Sudha Gopalakrishnan Brain Centre. The research centre housed at IIT-Madras has taken up the ambitious project of high-resolution brain imaging at the cellular level and the team, headed by professor Mohanasankar Sivaprakasam, is collecting and analysing big data sets running into petabytes. They are currently supported by generous donations from the likes of Infosys co-founder Kris Gopalakrishnan and corporate partnerships for high-end computing infrastructure.
Sivaprakasam, who also heads IIT-Madras’ Healthcare Technology Innovation Centre told businessline that while he was not aware of the specifics of the government’s plans to subsidise compute, any support in meeting this large cost is a welcome move.
“Today, GPUs are over priced because of a near monopoly situation and not enough competition in the market when it comes to GPU makers,” he said. “If you want to undertake large-scale and critical scientific research in India, you will also need to reach a stage when this infrastructure is available here locally,” he added.
As announced by minister Ashwini Vaishnav recently, the government is set to democratise compute facility for researchers, companies, and academia. “The cost per GPU hour is approximately ₹115.85, significantly lower than the global benchmark of $2.5–$3 per hour,” the statement from the IT ministry said. “High-end computing will be available at ₹150 per hour, with a 40% government subsidy reducing the cost to less than ₹100 per hour for common Compute access,” it added.
It’s not just the IITs but other universities too that are engaged in core AI research that could benefit from
Revathi Venkataraman, Chairperson, School of Computing, SRM Institute of Science and Technology says that their department is working on Genomic analysis using Deep Learning, Medical image analysis using Convolutional Neural Networks (CNNs) and other such research that involves analysis of huge data volumes and natural language processing. “While the initiative is in its early stages, academic institutions will likely benefit from lower-cost AI infrastructure, though formal application criteria remain unclear,” Venkataraman says. She also believes that the country should aim to develop GPUs domestically to achieve self-sufficiency in the sector.
Chiranjib Bhattacharyya, Professor and Chair, Department of Computer Science and Automation (CSA), Indian Institute of Science (IISC) points out that AI infrastructure is not just needed for undertaking research but also to train next-generation researchers. “High cost of GPUs and their limited availability is a significant infrastructural challenge currently faced by academia. We have high hopes that the authorities will prioritise academia and make the GPUs available to us. If that indeed materialises, then it would be a gamechanger,” Researchers at IISC plan to apply for compute once the rules are formalised.
Research teams in the corporate world too are enthused by plans to improve compute access.
Take the AI research team at Fractal Analytics that relies on large compute capacity for its various projects. The team’s text-to-image diffusion model Kalaido.ai trained on a public dataset of 70 million images and is capable of generating images from text prompts in English and 17 Indian languages. Similarly its healthcare AI platform Vaidya.ai was also trained on a curated dataset comprising over 850,000 plus images and text entries. “Currently foreign companies and researchers have the advantage of GPUs priced in dollars and Indian teams have the additional forex hit on GPUs. We hope we can move more of our workloads to India with this [announcement],” Suraj Amonkar, Chief AI Research and Platforms Officer at Fractal, said.
Uday Reddy Bondhugula, Professor, IISC and CTO of FSID-incubated Polymage Lab said that globally academia lags behind industry in AI model development and insufficient compute is a key reason. There are some notable exceptions of academia breakthroughs like MLIR compiler infrastructure, the Triton programming framework, and FlashAttention, among others, he adds.
However, researchers point out the challenges that still exist for academia to spearhead AI innovation.
“There is a shortage of skilled AI researchers and privacy concerns around IndiaAI datasets create hurdles. While BharatGPT and AI4Bharat involve academia, large industry-academia partnerships are still rare,” SRM’s Venkataraman said.
Availability of large volumes of data for training models is also a challenge we are yet to solve for, IISC’s Bhattacharyya noted. “A large number of young students must embrace research careers. This will strengthen India’s position as a key leader in the field of engineering of AI systems. For access to data, it will be important to collaborate with relevant industries,” he said.