It’s often a challenge for programmers and scientists to get time on high-performance supercomputers. These machines are expensive to build and maintain, but there’s no substitute for the massively parallel computing environment of a supercomputer. A new project at the Los Alamos National Laboratory’s High Performance Computing Division seeks to make supercomputers more accessible with a little help from some Raspberry Pi clusters.
Los Alamos National Laboratory (LANL) is home to several of the world’s most powerful supercomputers, including Trinity. That machine cost nearly $ 200 million to build, and its Intel Xeon Phi CPU cores are both powerful and power-hungry. Still, scientists need that sort of power for certain applications. For testing and running simpler programs, the modest ARM chips in the Raspberry Pi could be sufficient when you get enough of them together. LANL worked with Australian BitScope Designs to create its new Pi-powered supercomputer from 750 individual mini-computers.
The device is based on five rack-mount BitScope Cluster Modules. Each one has 150 Raspberry Pi 3 nodes networked together (that’s 750 total Pis). Each Raspberry Pi 3 has a Broadcom BCM2837 system-on-a-chip (SoC) with four 64-bit CPU cores clocked at 1.2GHz. They’re ARM Cortex-A53 reference cores, which are the same thing you’ll find in many budget smartphones running Qualcomm and MediaTek SoCs. This adds up to 3,000 available CPU cores for the full system, but it uses only a fraction of the power needed for a computer like Trinity. LANL estimates the system will need just 1,000 watts at idle and 2,000 watts during typical usage. The maximum load is 4,000 watts. Other supercomputers use between 10 and 25 megawatts of power.
Close-up of BitScope racks.
The Raspberry Pi-based supercomputer will be much slower than a “real” supercomputer, but the system architecture is similar to those more expensive systems. LANL envisions researchers testing their code on the BitScope system before porting the framework to a more powerful system that has a waiting list. Not only does this free up time on supercomputers for more important work, but it also costs much less for researchers to test code on the slower ARM-based systems.
BitScope plans to make the Cluster Modules available for purchase early next year. A single rack with 150 Raspberry Pi nodes will cost around $ 18,000-20,000. That works out to $ 120 per node. Of course, a Raspberry Pi board costs just $ 35 at retail, but these will be pre-configured and networked together for instant parallel computing. That’s not bad when you consider even smaller supercomputers running Intel and AMD chips could cost several million dollars.
Let’s block ads! (Why?)
ExtremeTechExtreme – ExtremeTech