The Tianhe and Taihu Light supercomputers, built with billions of dollars in investment, have allowed China to dominate the supercomputer rankings for a long time. However, instead of spending huge amounts of money on supercomputers, the U.S. National Laboratories have chosen a more cost-effective method by creating a supercomputer from a cluster of 750 Raspberry Pi development boards, successfully creating a more economical supercomputer. In the future, they plan to expand to 10,000 nodes to challenge the top 100 Peta-scale supercomputers globally. This HPC system has already begun testing and will be used for scientific research. The Raspberry Pi supercomputer was also showcased at the 2017 Supercomputing Conference held in Denver, USA this week.
This system consists of 5 modules, each containing 150 Raspberry Pi 3 B units in a 6U chassis. Since the Raspberry Pi 3 B is equipped with a 64-bit quad-core ARMv8 1.2GHz processor, having up to 750 Raspberry Pi 3 B units means there are over 3000 cores available, sufficient to meet the multi-core requirements of specific supercomputers.
It is worth mentioning that the domestic price of the Raspberry Pi 3 Model B is around 240 yuan, so the supercomputer made up of 750 Raspberry Pis costs about 180,000 yuan, which is already the lowest price for a supercomputer with over 3000 cores.
Using Raspberry Pi to build supercomputers is not a new concept. The Raspberry Pi, which is the size of a matchbox and has the functionality of a mini-computer, became very popular after its launch in 2012, even sparking a maker movement. A Raspberry Pi development board, available for around 1000 yuan, has become the first choice for many startups to develop and validate products. This DIY trend has even spread to the scientific community, with some using it for scientific research. For example, in 2012, Joshua Kiepert, a PhD student at the University of Arkansas, created a computing cluster using 32 Raspberry Pis at a cost of less than $2,000.
Later, a computer engineer from the University of Southampton in the UK used 64 Raspberry Pis and Lego blocks to create the Iridis-Pi supercomputer, costing less than 130,000 TWD. The Raspberry Pi supercomputer built by the Los Alamos National Laboratory, one of the U.S. Department of Energy’s two national laboratories, used ten times the number of Raspberry Pis, totaling 750 development boards.
Building a World-Class Supercomputer with 750 Raspberry PisThe supercomputer built by the Los Alamos National Laboratory mainly uses modular blade server cabinets provided by BitScope. Since it is entirely designed using Raspberry Pi development boards, it is much cheaper than traditional supercomputers and is also more energy-efficient, with each computing node averaging only 2 to 3 watts of power consumption.
According to Gary Grider, head of high-performance computing at Los Alamos National Laboratory, this Raspberry Pi supercomputer system has a total of 750 CPU computing nodes (using a 1.2 GHz Quad Core ARMv8 processor), with up to 3,000 cores available for computation, already possessing the computing power equivalent to a supercomputer. The laboratory plans to expand to challenge 10,000, or even up to 50,000 nodes in the future. In terms of computational performance, theoretically, it will be comparable to their Grizzly supercomputer system (ranked 85th globally with a performance of 1.5 PFLOPS), and even faster than Taiwan’s National Center for High-Performance Computing’s self-developed Peta supercomputer (ranked 95th globally with a performance of 1.325 PFLOPS).
Not only is the goal to break into the top 100, but the construction cost of this Raspberry Pi supercomputer is much cheaper than that of traditional supercomputers. Although Gary Grider did not disclose the actual construction cost, if calculated based on their use of the Raspberry Pi 3 Model B (priced at $35), the total cost for 750 Raspberry Pis is about $26,000 (approximately 780,000 TWD). Even with additional foundational construction costs, it is still far lower than the typical hundreds of millions of dollars spent on traditional supercomputers.
Gary Grider stated that developers can use Raspberry Pi modules and software to write applications on the supercomputer without spending billions of dollars on high-cost hardware and software to build an HPC test platform, which also consumes millions of watts (MW) of power to operate normally. Grider also revealed that their self-built Raspberry Pi supercomputer consumes an average of only 2 to 3 watts per computing node. Even when all 750 nodes are summed up, the power consumption is still far lower than that of many current supercomputers, saving both money and electricity, and the system can also be flexibly expanded as needed.
The Raspberry Pi supercomputer chassis consists of 5 blade modules, each capable of housing 150 Raspberry Pi development boards. The above image shows one of the blade modules, which is connected via Ethernet to establish a computing cluster. The chassis also integrates a switch and is equipped with a 48-volt power supply device for power.
However, creating a Peta-scale supercomputer using thousands to tens of thousands of Raspberry Pis is not an easy task. As the scale of the computing cluster increases, the connections and transmissions between different computing nodes become more complex, leading to potential access delays and insufficient transmission bandwidth, which can affect system performance. This is also a challenge for system expansion. Even Los Alamos National Laboratory initially could only expand to a maximum of 40 computing nodes. Later, they collaborated with HPC vendors BitScope and SICORP to redesign the chassis and integrate the system to solve these challenges, increasing from the original 40 nodes to 144, and now to 750 nodes, with plans to challenge thousands to tens of thousands of nodes in the future.
According to the SICORP website, the Raspberry Pi supercomputer is built using modular blade cabinets provided by BitScope. The chassis consists of 5 blade modules, each capable of housing up to 150 Raspberry Pi development boards. These computing modules can be connected via Ethernet to establish a computing cluster. The chassis also integrates a network switch and is equipped with a 48-volt power supply device for power.
However, Gary Grider emphasized that this Raspberry Pi supercomputer is not intended to replace other supercomputers but is hoped to provide developers and researchers with a new HPC computing platform for research testing and validation that not only possesses high-performance computing capabilities but is also more affordable. He also stated that in the future, in addition to continuously increasing the number of nodes to improve computing performance, they will also establish new network architectures and incorporate algorithms for bootstrapping, management, and disaster recovery, as well as focus on system and distributed storage research.
Submission: [email protected]
