Feature In the decade since India launched its National Supercomputing Mission (NSM), the nation has commissioned 37 machines with a combined power of 39 petaFLOPS, with another 35-petaFLOPS hybrid due to come online later this year. But while plenty of those machines use locally developed technology, India is yet to deliver on its ambition to become a leader or major semiconductor player.
India launched the NSM in 2015 to ensure the country has the computing power its government and research sector need.
The NSM has mostly achieved the latter goal. More than 13,000 scholars can access the NSM’s fleet, utilization rates sit at 85 to 95 percent, more than ten million workloads have run, and the fleet has helped to produce in excess of 1,500 peer-reviewed papers.
India’s supercomputing efforts are overseen by the Centre for Development of Advanced Computing (C-DAC), which in 2020 extended the nation’s supercomputing ambitions by calling for the creation of locally developed exascale chip designs, design and manufacture of exascale server boards, exascale interconnects, and storage including silicon photonics.
C-DAC senior director Sanjay Wandhekar tells The Register that effort is succeeding. “Our indigenization efforts have reached more than 50 percent, counting server nodes, interconnects, and system software,” he said.
Those nodes use a server design called “Rudra” that was developed under the NSM and powers a third of the program’s fleet.
The latest Rudra spec allows use of recent GPUs from Nvidia and AMD, plus CPUs from Intel and AMD. India’s homegrown processor, known as “AUM,” remains at least two years from completion.
While India waits for its homespun CPU, tech giants are happy to help the NSM. Intel’s Xeon processors currently dominate the NSM fleet. However, AMD’s corporate VP and managing director for India, Vinay Sinha, tells The Register AMD has co-designed Rudra servers with C-DAC and aligned its roadmap for EPYC processors and Instinct accelerators with the NSM.
Red Hat’s general manager for India and South Asia, Navtez Bal, says RHEL forms the base OS on many NSM clusters, but the entire user-facing stack – job schedulers, resource managers, monitoring – uses C-DAC’s customized open source, eliminating commercial lock-in. CDAC told us it uses CentOS, which draws from RHEL, on some NSM machines.
Every new Rudra installation includes direct liquid cooling developed by C-DAC. Wandhekar says it delivers a 10 percent node-level power saving under full load and contributes to a site-wide PUE of 1.20-1.25 – roughly a 20 percent improvement over air-cooled equivalents. Pricing is at par with product from global OEMs, but direct component procurement and local manufacturing shave at least 15 percent off costs.
The cost of Rudra servers “matches other vendors,” Wandhekar claims, noting the fact C-DAC assembles servers in house gives it a “15 percent price advantage.” C-DAC has also developed an interconnect called Trinetra and uses it in three NSM machines. More on that later.
How indigenous is it really?
While C-DAC claims NSM machines are 50 percent indigenous, others dispute that figure.
“Definitely less than 40-50 percent,” counters Professor Rupesh Nasre of the Indian Institute of Technology (IIT) Madras. “Reaching 50 percent would itself be a big achievement.”
IIT Bombay’s Professor Madhav Desai points to the missing ecosystem that turns hardware and software prototypes into reliable products. “The objective remains to create a balanced ecosystem that promotes indigenous innovation while engaging with global expertise.”
Analysts echo the trade-offs. “The hybrid approach accelerates capability, but reliance on foreign IP exposes us to export controls and supply chain volatility,” says Sushovan Mukhopadhyay, director analyst at Gartner.
“India brings market scale and design talent, but the technology core still rests with global players,” notes Sharath Srinivasamurthy, associate vice president at IDC India.
“Fab access is the single biggest bottleneck. AUM needs TSMC 5nm – until India has its own advanced nodes, sovereignty will stay partial,” claims Gareth Owen, director at Counterpoint Research.
India’s AI vision calls for 80 exaFLOPS of infrastructure
India turns on a new supercomputer
India flips the switch on three homebrew supercomputers
India plans 10,000-GPU sovereign AI supercomputer
Even if India can deliver AUM in a form suitable for supercomputers, the nation lacks other technologies.
C-DAC’s Trinetra interconnect, for example, produces latency more than double that offered by commercial InfiniBand products.
India is also nowhere near producing an indigenous GPU.
Indian Institute of Science (IISc) Professor Sathish Vadhiyar leads a research effort to create “a comprehensive library of indigenous IP cores” for GPUs, including neuromorphic accelerators and optical interconnects. But the project has not set a delivery date.
NSM 2.0
While India strives to develop more of its own supercomputing stack, users are enjoying the current NSM fleet.
“The community is very happy,” Wandhekar says. “Previously, users faced fragmentation and spent significant time just buying and configuring systems. Now NSM provides complete solutions.” Training has reached 26,000 people through nodal-center workshops.
At the Indian Institute of Tropical Meteorology in Pune, a hub for climate R&D from short-range forecasts to decadal projections, Dr Murali Krishna says the payoff is operational:
“We’ve handed the Bharat Forecast System (~6 km global) and sub-seasonal-to-seasonal models to IMD [India Meteorological Department] for daily use.”
IIT Kharagpur director Prof Suman Chakraborty tells us: “We began by importing muscle. Today we are building the brain, the spine, and the heartbeat ourselves.”
The current phase of the NSM ends on December 31. C-DAC has already determined the goals for NSM 2.0: delivering 1.5 exaFLOPS worth of systems within five years of the program’s approval.
The initial wave will “likely deploy 100-petaFLOPS systems” on Intel/AMD hardware. C-DAC’s plan calls for a second wave of exascale flagships to run on the indigenous AUM processor, based on an Arm Neoverse V2 design developed with MosChip and Socionext and built on TSMC’s 5nm node. C-DAC expects AUM to arrive around 2027, and be ready for deployment from 2028.
For C-DAC, success is not just raw capacity. “We aim for performance that is meaningful for India-specific challenges,” Wandhekar says. The top non-hardware gaps are “skilled people to design, operate, and optimize these systems and a mature domestic manufacturing ecosystem.” Talent retention is difficult, though Wandhekar notes many stay in the broader ecosystem for several years.
Prof Vadhiyar calls for a multi-pronged push: “Large exascale-class systems with full indigenization, continued R&D in hardware, software, and applications.”
The final challenge is demand. Prof Chakraborty says India needs “powerful mechanisms for demand creation,” proposing a “national HPC marketplace” where industry can upload problems and academia can build solutions. “We need plug-and-play HPC for MSMEs [micro, small, and medium enterprises]. Until we bridge this gap, our machines will remain more capable than they are impactful.” ®
