Fermi (supercomputer)
| Company type | Supercomputer |
|---|---|
| Industry | High-performance computing |
| Founded | June 2012 |
| Headquarters | Bologna , Italy |
Area served | Scientific research |
| Products | High-performance computing services |
| Owner | CINECA |
Fermi is a supercomputer delivering 2.097 petaFLOPS of performance, housed at CINECA.
History
Acquired in June 2012 and fully operational by August of the same year, FERMI stands as CINECA's flagship high-performance computing system. It functions as Italy’s national tier-0 supercomputer dedicated to scientific research and is part of the European HPC network (PRACE). The acquisition was funded by the Italian Ministry of Education, Universities and Research.
In the June 2012 edition of the TOP500 ranking, Fermi secured the seventh position among the fastest supercomputers globally.
Additionally, on the Graph500 list, it achieved fifth place with a performance of 2,567 gigaTEPS (giga traversed edges per second).
Specifications
FERMI is based on IBM’s final iteration of the petascale supercomputer series, the Blue Gene/Q. The system consists of 10 racks, each containing two midplanes, which together provide a total of 10,240 compute nodes and 163,840 processing cores.
- Every compute node is equipped with a 1.6 GHz IBM processor featuring 16 A2 cores, 16 GB of RAM, and network capabilities. A node card comprises 32 compute nodes; sixteen such node cards form one midplane. Each rack includes two midplanes plus two I/O drawers, resulting in 16,384 cores per rack. Compute nodes operate using a lightweight Linux-like kernel known as CNK.
- The compute nodes do not have local disk storage; instead, dedicated I/O nodes manage input/output operations.
- Users access the system via ssh through front-end login nodes at login.fermi.cineca.it, which run RedHat Linux version 6.2. Parallel applications must be cross-compiled on these front-end nodes before execution on allocated compute partitions.
CINECA’s setup features:
- Two racks equipped with 16 I/O nodes each (minimum job allocation: 64 nodes / 1,024 cores).
- Eight racks containing 8 I/O nodes each (minimum job allocation: 128 nodes / 2,048 cores).