PARAM

PARAM
Formation	late 1980's
Type	Supercomputers
Parent organization	Centre for Development of Advanced Computing (C-DAC)

PARAM is a series of Indian supercomputers designed and assembled by the Centre for Development of Advanced Computing (C-DAC) in Pune. PARAM means "supreme" in the Sanskrit language, whilst also creating an acronym for "PARAllel Machine". As of November 2022 the fastest machine in the series is the PARAM Siddhi AI which ranks 163rd in world, with an Rpeak of 5.267 petaflops.

History

C-DAC was created in November 1987, originally as the Centre for Development of Advanced Computing Technology (C-DACT). This was in response to issues purchasing supercomputers from foreign sources. The Indian Government decided to try and develop indigenous computing technology.

PARAM 8000

The PARAM 8000 was the first machine in the series and was built from scratch. A prototype was benchmarked at the "1990 Zurich Super-computing Show": of the machines that ran at the show it came second only to one from the United States.

A 64-node machine was delivered in August 1991. Each node used Inmos T800/T805 transputers. A 256-node machine had a theoretical performance of 1GFLOPS, however in practice had a sustained performance of 100-200MFLOPS. PARAM 8000 was a distributed memory MIMD architecture with a reconfigurable interconnection network.

The PARAM 8000 was noted to be 28 times more powerful than the Cray X-MP that the government originally requested, for the same $10 million cost quoted for it.

Exports

The computer was a success and was exported to Germany, United Kingdom and Russia. Apart from taking over the home market, PARAM attracted 14 other buyers with its relatively low price tag of $350,000.

The computer was also exported to the ICAD Moscow in 1991 under Russian collaboration.

PARAM 8600

PARAM 8600 was an improvement over PARAM 8000. In 1992 C-DAC realised its machines were underpowered and wished to integrate the newly released Intel i860 processor. Each node was created with one i860 and four Inmos T800 transputers. The same PARAS programming environment was used for both the PARAM 8000 and 8600; this meant that programs were portable. Each 8600 cluster was noted to be as powerful as 4 PARAM 8000 clusters.

PARAM 9000

The PARAM (param vashisht lega) 9000 was designed to be merge cluster processing and massively parallel processing computing workloads. It was first demonstrated in 1994. The design was changed to be modular so that newer processors could be easily accommodated. Typically a system used 32–40 processors, however it could be scaled up to 200 CPUs using the clos network topology. The PARAM 9000/SS was the SuperSPARC II processor variant, the PARAM 9000/US used the UltraSPARC processor, and the PARAM 9000/AA used the DEC Alpha.

PARAM 10000

The PARAM 10000 was unveiled in 1998 as part of C-DAC's second mission. PARAM 10000 used several independent nodes, each based on the Sun Enterprise 250 server; each such server contained two 400Mhz UltraSPARC II processors. The base configuration had three compute nodes and a server node. The peak speed of this base system was 6.4 GFLOPS. A typical system would contain 160 CPUs and be capable of 100 GFLOPS But, it was easily scalable to the TFLOP range. Exported to Russia and Singapore.

Further computers

Further computers were made in the PARAM series as one-off supercomputers, rather than serial production machines. From the late 2010s many machines were created as part of the National Supercomputing Mission.

Supercomputer summary

PARAM Timeline
Name	Release Year	Notes	Rmax	Rpeak	Location
PARAM 8000	1991	Inmos T800 Transputers, Distributed Memory MIMD, 64 processors			Multiple
PARAM 8600	1992	Intel i860, 256 processors	5 GFLOPS		Multiple
PARAM 9900	1994	clos network. SuperSPARC II, UltraSPARC and DEC Alpha variants, 32 to 200 processors			Multiple
PARAM 10000	1998	Sun Enterprise 250, 400Mhz UltraSPARC UltraSPARC II processor, 160 processors	6.4 GFLOPS
PARAM Padma	2002	1TB storage, 248 IBM Power4 – 1 GHz, IBM AIX 5.1L, PARAMNet. PARAM Padma was the first Indian machine ranked on a worldwide supercomputer list.	1024 GFLOPS
PARAM Yuva	2008	4608 cores, Intel 73XX – 2.9 GHz, 25 to 200 TB, PARAMnet 3.	38.1 TFLOPS	54 TFLOPS
PARAM Yuva II	2013	Created in three months at a cost of ₹160 million (US$2 million) - first Indian supercomputer to achieve more than 500 teraflops.	360.8 TFLOPS	524 TFLOPS
PARAM Kanchenjunga	2016		15 TFLOPS
PARAM Bio-Embryo			100 TFLOPS		C-DAC Pune
PARAM Bio-Inferno			147.5 TFLOPS		C-DAC Pune
PARAM Shrestha			100 TFLOPS		C-DAC Pune
PARAM Rudra		based on Intel Xeon 2nd Generation Cascade Lake dual socket Processors	138 TFLOPS		C-DAC Pune
PARAM Neel		India's first HPC system that uses the Fujitsu A64fx- NSP1 CPU, an ARM processor with 48 cores and a speed of 1.8 GHz	100 TFLOPS		C-DAC Pune
PARAM Shivay	2019	192 CPU compute nodes, 20 High memory nodes, 11 GPU compute nodes, Cost Rs 32.5 crore	0.43 PFLOPS	0.84 PFLOPS	Indian Institute of Technology (BHU) Varanasi
PARAM Brahma	2019	1PB storage	0.85 PFLOPS	1.7 PFLOPS	IISER Pune
PARAM Siddhi-AI	2020	Nvidia DGX SuperPOD based networking architecture, HPC-AI engine software frame works and cloud platform from C-DAC	4.6 PFLOPS	5.267 PFLOPS	C-DAC Pune
PARAM Sanganak	2020			1.67 PFLOPS	IIT Kanpur
PARAM Yukti				1.8 PFLOPS	JNCASR, Bengaluru
PARAM Utkarsh	2021	Based on Intel Cascade Lake processor and NVIDIA Tesla V100 GPU with 100Gbps infiniband non-blocking interconnect	838 TFLOPS		C-DAC Bengaluru
PARAM Smriti	2021			838 TFLOPS	NAIB Mohali
PARAM Seva	2021	based on heterogeneous and hybrid configuration of Intel Xeon Cascade lake processors, and NVIDIA Tesla V100.		838 TFLOPS	IIT Hydrabad
PARAM Spoorthi	2021			100 TFLOPS	SETS, Chennai
PARAM Pravega	2022	It runs on CentOS 7.x, has 4 petabytes of storage, 3.3 PFLOPS		3.3 PFLOPS	Indian Institute of Science, Bengaluru
PARAM Ganga	2022			1.67 PFLOPS	IIT Roorkee
PARAM Shakti	2022		850 TFLOPS	1.66 PFLOPS	IIT Kharagpur
PARAM Ananta	2022			838 TFLOPS	IIT Gandhinagar
PARAM Himalaya	2022			838 TFLOPS	IIT, Mandi
PARAM KAMRUPA	2022	107 CPU nodes, 10 GPU nodes, 9 high memory nodes, 740 CPU cores, 102400 CUDA cores. It runs on low and high microwave power with active and passive high energy source.	838 TFLOPS	1.5 PFLOPS	Indian Institute of Technology Guwahati
PARAM Porul	2022	107 CPU nodes, 10 GPU nodes, 39 high memory nodes, 102400 CUDA cores.		838 TFLOPS	National Institute of Technology, Tiruchirappalli

PARAMNet

PARAMNet is a high speed high bandwidth low latency network developed for the PARAM series. The original PARAMNet used an 8 port cascadable non-blocking switch developed by C-DAC. Each port provided 400 Mb/s in both directions (thus 2x400 Mbit/s) as it was a full-duplex network. It was first used in PARAM 10000.

PARAMNet II, introduced with PARAM Padma, is capable of 2.5 Gbit/s while working full-duplex. It supports interfaces like Virtual Interface Architecture and Active messages. It uses 8 or 16 port SAN switches.

PARAMNet-3, used in PARAM Yuva and PARAM Yuva-II, is next generation high performance networking component for building supercomputing systems. PARAMNet-3 consists of tightly integrated hardware and software components. The hardware components consist of Network Interface Cards (NIC) based on CDAC's fourth generation communication co-processor "GEMINI", and modular 48-port Packet Routing Switch "ANVAY". The software component "KSHIPRA" is a lightweight protocol stack designed to exploit capabilities of hardware and to provide industry standard interfaces to the applications. Other application areas identified for deployment of PARAMNet-3 are storage and database applications.

Operators

PARAM supercomputers are used by both public and private operators for various purposes. As of 2008, 52 PARAMs have been deployed. Of these, 8 are located in Russia, Singapore, Germany and Canada. PARAMs have also been sold to Tanzania, Armenia, Saudi Arabia, Singapore, Ghana, Myanmar, Nepal, Kazakhstan, Uzbekistan, and Vietnam.