Daily Image

Click here or on the picture for a full size image.

13 TFLOPS at 2400 Watt

Submitter: John Romein
Description: This picture shows the latest acquisition of the DAS-4 experimental computer cluster. This system hosts eight of the fastest Graphical Processing Units (GPUs), one of which is shown here. It is absolutely a number-crunching monster.

Since a few years, GPUs are widely used for general-purpose computing. They are much more efficient in performing data-parallel computations than general-purpose processors; many applications run tens of times faster on a single GPU than on a CPU, thus imagine the speed of eight GPUs. The speed increase is obtained by exploiting massive parallelism: each of these GPUs hosts 512 cores, and each core works on tens of tasks simultaneously, closely cooperating with neighbouring cores. In total, the system can work on 131,096 tasks concurrently, 131,072 on the GPUs and 24 on the CPUs.

Historically, GPUs were hard to program, since the hardware was designed to support graphical operations only (e.g., the hardware could not write the result of a computation to a random location in memory). The past decade, GPUs have adopted more and more features of CPUs and have become much more suitable for general-purpose programming. Emerging programming languages like CUDA and OpenCL make it much easier to develop high-performance computing applications for GPUs than a pixel-rendering language like OpenGL.

The system shown has a theoretical peak performance of 13 TFLOPs, but most applications achieve some 10%-40% of the theoretical peak, due to memory bandwidth limitations and the relatively slow PCIe bus that connects the card with the rest of the system. The two orange fans are mounted in front of a ribbed heat exchanger. The fans on the right blow cold air through the system at tornado speeds to get rid of some 2400 Watts of heat.

The system will be used to do GPU-scaling performance experiments and to do "green computing" research. If you are interested in using this machine for your research, then drop me a line. But your application had better exhibit a high degree of parallelism!

Copyright: ASTRON
  Follow us on Twitter
Please feel free to submit an image using the Submit page.