The PARAM Supercomputers, From C-DAC's PARAM 8000 in 1991 to a National Supercomputing Mission With 37 Systems, About 40 Petaflops and Rudra Servers Made in India

On 10 December 2025 the government told Lok Sabha that the National Supercomputing Mission had deployed 37 supercomputers with a combined 40 petaflops, at institutions that include IISc, the IITs, C-DAC, research labs and several Tier-II and Tier-III cities. A petaflop is a quadrillion (a thousand trillion) arithmetic operations a second. Most of the named machines in that list carry the name PARAM, the line that C-DAC, the Centre for Development of Advanced Computing, began building in 1988 and first delivered in 1991.

What the mission reports, and three slightly different tallies

The National Supercomputing Mission (NSM) was launched in April 2015 with an outlay of Rs 4,500 crore. The Ministry of Electronics and IT (MeitY) and the Department of Science and Technology run it jointly, and C-DAC in Pune and the Indian Institute of Science in Bengaluru carry it out. According to the 10 December reply, the systems have served more than 13,000 researchers, among them more than 1,700 PhD scholars from over 260 institutions, and have completed more than one crore (ten million) compute jobs. They are credited with more than 1,500 research papers.

The official tally shifts by a few systems depending on the ministry and the date. On 3 December 2025 the Ministry of Science and Technology told Parliament that 37 systems with about 39 petaflops had been designed, built and commissioned. MeitY’s reply a week later gave 37 systems and 40 petaflops. On 8 January 2026, at the inauguration of PARAM Rudra at IIT Bombay, the NSM’s mission director said the total had reached 38 systems and 44 petaflops.

The plan was a grid from the start. A PIB note of 23 March 2020 says the 2015 approval envisaged a network totalling 15 to 20 petaflops, later revised to 45, including three systems of at least 3 petaflops each, across institutions of national importance by 2022. The National Knowledge Network is named as the backbone that links them. The January 2026 figure of 44 petaflops sits against that revised 45.

Why C-DAC was set up in 1988

C-DAC’s own About Us page gives the origin in two sentences. According to it, C-DAC was set up in 1988 to build supercomputers, in the context of the United States denying India the import of them, and that the PARAM line began with a machine of 1 gigaflop that year. A gigaflop is a billion operations a second. That page gives neither the date of the denial nor the Cray model involved, and none of the sources opened for this article do either, so both are left out here. Popular retellings quote higher benchmark speeds for the first PARAM, and this article uses only the 1 GF on C-DAC’s page.

A 2006 profile in Voice&Data magazine says the United States had refused to let Cray sell its supercomputer to India. Dr Vijay Bhatkar is credited there with leading the C-DAC team that produced India’s first supercomputer, PARAM, in 1991, and the profile records that he received the Padma Shri in 2000.

That first machine was PARAM 8000, a parallel computer, which means many processor chips share one problem. C-DAC’s case study of its work with Moscow says a PARAM 8000 was first installed at the Institute of Computer Aided Design of the Russian Academy of Sciences in 1991 to 1992, built on transputer chips, and that the installation held 128 nodes. Fluid mechanics and structural analysis software was parallelised on it. C-DAC’s page traces the collaboration to a science programme signed in Moscow in June 1987.

PARAM Yuva II and the TOP500 list

Twice a year, the TOP500 ranks the 500 fastest computers by their speed on a standard maths test called Linpack. Each entry carries a measured speed, Rmax, and a theoretical top speed, Rpeak. Ranks fall as faster machines arrive, which is why every rank below carries the date of its list.

In February 2013 C-DAC unveiled PARAM Yuva II at Pune. Voice&Data’s report of 9 February says peak compute rose from 54 teraflops on the earlier PARAM Yuva to 524, and that the machine sustained 360.8 teraflops on Linpack. A teraflop is a trillion operations a second. C-DAC’s own comparison was that the result would have ranked 62nd on the November 2012 TOP500 and first in India, and it described the machine as India’s first 500-teraflop supercomputer.

PARAM Yuva II first appeared on the real list in June 2013, at rank 70, with a measured 386.71 teraflops. By November 2013 it was at 84, and by November 2016 at 448.

Pratyush and Mihir, then Arka and Arunika

Pratyush and Mihir are not PARAM machines and do not appear in the NSM’s list of systems. They are Cray XC40 machines of the Ministry of Earth Sciences. IITM Pune’s own account says the Union minister for earth sciences dedicated Pratyush at the Indian Institute of Tropical Meteorology in Pune on 8 January 2018 and calls it India’s first multi-petaflop supercomputer. Mihir went to the National Centre for Medium Range Weather Forecasting (NCMRWF) in Noida.

TOP500 gives Pratyush 119,232 cores, a measured 3.76 petaflops and a theoretical 4.01. Mihir has 83,592 cores, 2.57 measured and 2.81 theoretical. The two theoretical figures add up to the 6.8 petaflops that the ministry’s 2024 release gives as its capacity before Arka and Arunika. Pratyush entered the TOP500 at rank 39 on the June 2018 list and Mihir at 66. On the November 2020 list they stood at 77 and 144.

Their successors came on 26 September 2024, when the Prime Minister inaugurated Arka at IITM and Arunika at NCMRWF. Per the ministry’s release, the project is worth Rs 850 crore, with 11.77 petaflops at IITM, 8.24 petaflops at NCMRWF and a separate 1.9-petaflop system for AI and machine learning. Total ministry capacity rises to 22 petaflops from 6.8. According to the release, the models these machines run will reach resolutions of 1 km or finer over selected Indian regions and should sharpen forecasts of cyclones, heavy rain, hailstorms, heat waves and droughts. Why the monsoon is so hard to forecast is covered in an earlier tuput explainer. On the June 2026 TOP500 list, Arka ranks 219 at 8.28 measured petaflops and Arunika 274 at 5.94.

PARAM Siddhi-AI and the AI facility

PARAM Siddhi is the mission’s AI machine, installed in 2020 at the National AI Facility at C-DAC Pune. In the government’s list it is given as 6.5 petaflops and 210 AI petaflops. Voice&Data’s report of 6 October 2020 explains the pair: 6.5 petaflops is the peak in double precision, the high-accuracy arithmetic of conventional science codes, and 210 is for AI work. According to that report the system is built from 42 NVIDIA DGX A100 units on NVIDIA’s SuperPOD design with HDR InfiniBand networking, runs C-DAC’s own HPC-AI software, and was conceived and architected by C-DAC’s Abhishek Das. C-DAC named healthcare, education, energy, cyber security, space, automotive and agriculture as target areas.

TOP500 lists PARAM Siddhi-AI at 41,664 cores, a measured 4.62 petaflops and a theoretical 5.27. PARAM Siddhi-AI entered at rank 62 on the November 2020 list, then stood at 89 in June 2021, 102 in November 2021, 120 in November 2022 and 185 in June 2024, the last row on its TOP500 page. Its name is not among the Indian entries on the June 2026 list. A separate programme rents out GPUs for AI work, covered in the tuput piece on the IndiaAI Mission.

Rudra servers and who built what

PARAM Rudra is the mission’s line of indigenously built hardware. In the 10 December reply, the government says C-DAC designed the Rudra server board and transferred the technology to three Indian electronics manufacturers, and that a high-speed network between nodes, Trinetra, has been built and tested at 40 and 100 gigabits a second. Cooling technology and the full system software stack are also indigenous, the reply adds, and applications for flood forecasting, urban weather modelling, seismic data processing for oil and gas, genomics and materials science have been deployed with users such as the India Meteorological Department, ONGC, the Central Water Commission and the Central Pollution Control Board. PARAM Shavak, a desktop supercomputer, is made for colleges.

A week earlier, the 3 December reply had named the manufacturers, VVDN Technologies and Kaynes Technologies, and said 6,000 Rudra servers were deployed in PARAM Rudra supercomputers with 1,500 more being made. On 26 September 2024 the Prime Minister dedicated three PARAM Rudra systems worth around Rs 130 crore, at the Giant Metrewave Radio Telescope’s centre in Pune, IUAC in Delhi and the S.N. Bose Centre in Kolkata. In the same address he placed the mission next to the quantum mission, which tuput covered in a separate article. IIT Bombay’s 3-petaflop PARAM Rudra followed on 8 January 2026. On the June 2026 TOP500 list, C-DAC’s PARAM Rudra-104 ranks 245 at 7.03 measured petaflops.

Two people quoted by The Register in November 2025 measure the Indian share of the build differently. C-DAC’s Sanjay Wandhekar puts indigenisation above 50 percent, counting server nodes, interconnects and system software, while Professor Rupesh Nasre of IIT Madras puts it below 40 to 50 percent and adds that reaching 50 would itself be a big achievement. The same report says Trinetra, used in three NSM machines, has latency more than double that of commercial InfiniBand.

What the next stage aims at

Ten more supercomputers are being commissioned, the 3 December reply says, including a national facility with India’s largest HPC and AI set-up, at 20 petaflops. A week later, the 10 December reply put the work in progress at 6 new systems with an outlay of Rs 680 crore. Also on 3 December, the ministry said a roadmap called NSM 2.0 is being developed for the next generation of supercomputing in India, and that it contemplates India’s own processors and hardware with industry partners.

The Register reports that C-DAC’s goal for NSM 2.0 is 1.5 exaflops of systems within five years of the programme’s approval. An exaflop is 1,000 petaflops. The first wave would run on Intel and AMD hardware at around 100 petaflops a system, and a second wave would use AUM, an Arm-based processor C-DAC expects around 2027 for deployment from 2028. Counterpoint Research’s Gareth Owen told the magazine that access to chip fabrication is the biggest bottleneck, a subject tuput covers in its piece on India’s fab plans.

The June 2026 TOP500 list carries several Indian-sited systems. The highest is Shakti Cloud, a system at Yotta Data Services built on NVIDIA H100 GPUs, at rank 32 with 84.31 measured petaflops. Among the others, C-DAC’s AIRAWAT-PSAI ranks 211 at 8.50 petaflops, Arka 219, PARAM Rudra-104 245 and Arunika 274.