CEA
HPC Calcul haute performance
AccueilComplexe de calculLogiciels pour le HPCRecherche et developpementsCollaborations
 
Research and development / Summary

Research and development

The experience gained at the CEA computing centers over the last 10 years allows a certain number of observations to be made.

User needs are leading to a strong increase in demand for computing power and capacity to manage data. The development of physical models is leading to a strong increase in the flow of data, up to several tens of Terabytes per day, which are being transferred from the computer to the storage and visualization systems.

Moore's Law, which describes the technological development of microprocessors, no longer involves an increase in the operating frequency of processors; rather, it is being fulfilled by an increase in the number of cores per processor. The capacity of a computer unit remains almost constant; only by parallelism is the processing capacity of a machine being increased.

It can therefore be envisaged that future generations of supercomputer will integrate at least 100,000 computer cores, which is 10 times more than the current computer complexes.  Combined with an exceptional input/output system, this will lead to computers which are much more complex to design and control and, in all cases, exceed the natural evolution of machines in the marketplace.

Numerical simulation has always progressed through advances in three areas: computers, physical models and mathematical methods, the codes and the software. Advances made in the hardware have enabled great improvements in the other two factors. This could not be achieved without a continuous research effort in the fields of physics, mathematics and computer science.

Being aware of the issues raised by these observations, the CEA developed an R&D programme based on a technology survey carried out by experts from the computing complex and in collaboration with strategic participants from the market place such as computer manufacturers, processor manufacturers and "middleware" developers. The programme also considered the design of new architectures and programming models.


View of an Inca cluster (18 blades). Each Nehalem blade has dual processors (4 cores each), hence 8 cores per blade.

This equipment will form an integral part of the Tera 100 prototype which contains 24 Inca clusters.