SC12: HPC cluster with Nvidia mobile graphics and ARM SoCs

As part of the EU project, the company Montblanc Seco developed with Nvidia components for supercomputers with ARM SoCs. The Italian company E4 Computer Engineering now links eight to 24 blades with Qseven modules with Tegra 3 processors and GPGPU accelerators type Quadro 1000M at comparatively economical HPC clusters. After Nvidia Project "Carma" - CUDA for ARM - called the E4 systems Carma Carma microcluster or cluster.

The microcluster with eight blades, each one ARM SoC and GPU module carries, comes from a 520-watt power supply. E4 is the maximum computing power of 2.16 teraflops of single-precision floating-point calculations in (single precision). This corresponds to the performance of the eight Quadro 1000M graphics chips, each of which reach 96 CUDA cores of 270 Gflops. The CUDA 2.1 cores process only FP32 values, and no dual-precision numbers.

Nvidia sells a single of these blades as Carma devkit via Seco for around 630 euros and developer. E4 grabs again in Carma cluster together up to 12 nodes, equipped with Carma2 boards each node will have two Quadro 1000M, total come together to 6.48 teraflops SP computing power. E4 is setting up a 1.5-kW power supply. Carma the cluster can also be purchased exclusively with Tegra 3 processors, then a total of 48 of these SoCs stuck with four Cortex-A9 cores in it to 192 ARM cores. Sure enough, a 400-watt power supply. Are optional except for the Qseven board QuadMo747-X/T30
Tegra T30 with others available.

Where supported by the EU with around 8 million Euro project Montblanc as well as work with the Research Centre J├╝lich and the Leibniz Rechenzentrum Garching. The aim is efficient exascale supercomputers; 2017, a system with 200 PFlops stand on the Top500 list. For the first HPC prototypes Montblanc has now the ARM SoC Samsung Exynos selected 5 with two Cortex-A15 as it is at the SC12 conference was announced.

Nvidia is working under the Project Denver to SoCs with proprietary ARM cores, but also has a license for the Cortex-A15. This could come in the Tegra version Wayne used. The internally developed ARM cores could be either 64-bit versions, as well as AMD, STMicroelectronics, Calxeda, Applied Micro and others are planning - but in part based on the ARM Cortex-A57 designs and A53.

No comments:

Post a Comment