Skip to content

Barcelona have one of the world’s most efficient supercomputers based on ARM, can also become the fastest?

April 4, 2012

The National Supercomputing Center in Barcelona , Spain, is preparing to mount one of the most energy efficient supercomputers in the world, based on ARM processors. The most interesting thing is that over time could also become the fastest available, ahead of models advanced x86 processors such as Intel Xeon or AMD Opteron. Alex Ramirez, a center manager is responsible for building the monster computer, to be called Mont-Blanc, and used a version not “similar” to the chips used in devices such as tablets and smartphones, but exactly the same SoC available in many of these devices for personal entertainment.


Mont Blanc integrated NVIDIA Tegra 3 chips with ARM Cortex-A9 processors quad-core, used in products such as ASUS Prime Transformer . As of May, the team will begin assembling the first prototype of the supercomputer, whose main characteristic is the low energy consumption compared to other teams in the same category based on x86 architecture, most at the moment, or prior to RISC . Tegra chips 3 is responsible for communications between different parts of the system, and as for the graphic, will have low power multicore produced by NVIDIA, which could lead GeForce 520MX chip used in mobile systems inside the Optimus platform.


The center plans to test the first version of the super computer in June to gain a foothold in the Top 500 supercomputer list, which keeps a strict record of the performance of the top 500 supercomputing machines in the world, through the solution for Linpack benchmark. But the project’s main objective is to get one of the leading positions of the Green 500 list, which qualifies the equipment for energy efficiency rather than raw power. Ramirez wants to place Mont Blanc among the 10 most efficient machines on the planet. If plans continue its course assembly and the estimates are correct, the system would get a high efficiency of 7 Gigaflops per watt of consumption, equivalent to 7,000 million mathematical calculations per second per watt. The secret is to use graphics chips too low to support GPGPU to help General computing cores of many Tegra chips to be used, thus increasing overall performance and efficiency in consumption.


But could a system based on ARM to surpass the most powerful supercomputers x86? The first version of Mont-Blanc will get a spot in the Top 500 mounted between 2,000 and 4,000 sheets Tegrarunning in parallel. In the words of Alex Ramirez: “Instead of using a few high-performance processors are going to use many low-power processor and average yield.” With the ARM Cortex-A9 processors there is a “limit” in terms of power, but when the new NVIDIA Tegra 4 with core ARM Cortex-A15 64-bit to hit the market next year, performance will soar, reaching up to four times the performance per watt today, managing to keep a team like Mont Blanc in the list of more efficient machines – an imperative need today – while climbing positions is achieved in an extraordinary way in the demanding Top 500.


There is a small barrier to Mont-Blanc meets its objectives, the research software commonly used in this type of “clusters” and supercomputing machines is written for an entirely different architecture, which means that the Barcelona team have to worry about something rather than prepare the machine to pass the performance tests. Applications must be rewritten to leverage the power of both parallel ARM CPU as the GPU, and that’s not exactly easy task. From the center are quite sure they will be able to build the basis for the super computers of the future, be getting on top of the Top 500 list in 2017, five years into the future – assuming and hoping that the Maya are just a prank – The system could be, according to the project, around 200 Petaflops of computing power, as many as 20 times more potent than the current queen of supercomputing, Computer Nippon K.

Obviously the Spanish team’s estimates are based on their interpretation of the evolution of other architectures such as x86, something very relative, because at any moment may come a greater development than expected, leading to ruin the aspirations of the BSC CNS. At some point I said that ARM is gaining momentum almost unprecedented, and that at some point will exceed both commercially and technically to Intel and AMD, but this is out of any scale.