|
|
| Machine type |
RISC-based distributeded-memory multi-processor |
| Models |
FX1 |
| Operating system |
Solaris (Sun's Unix variant) |
| Connection structure |
Fat Tree |
| Compilers |
Parallel Fortran 90, OpenMP, C, C++ |
| Vendors information web page: |
— |
| Year of introduction |
2009 |
System parameters:
| Model |
FX1 |
| Clock cycle |
2.52 GHz |
| Theor. peak performance |
|
| Per core (64-bits) |
10.1 Gflop/s |
| Maximal |
— |
| Main memory |
|
| Memory/node |
≤ 32 GB |
| Memory/maximal |
— |
| Communication bandwidth |
|
| Point-to-point |
≥ 2 GB/s |
| Aggregate |
— |
Remarks
In April 2009 Fujitsu officially put a first FX1 HPC systems to work at JAXA
(Japanese Space Exploration Agency) in Japan. The information issued about the
system is extremely limited. For, instance there is no corresponding web page to
be found for the system at the Fujitsu website. This may be due to the fact that
Fujitsu has no plans yet to market the system outside Japan. Still, the machine
has some highly interesting features and it may become the base for the 10
Pflop/s Japanese successor of the Earth Simulator after Hitachi and NEC pulled
out of this project.In that case, however, the successor, the SPARC VIII will
be used.
From the scant information that is available, we know that the system is based
on the same SPARC VII processor that also is employed in Fujitsu's enterprise
servers like the M8000 and M9000. This quad-core processor (see the
SPARC processor) has a peak performance of just over 40
Gflop/s. The bandwidth from memory to the CPU is respectable: 40 GB/s, more than
in the Intel Nehalem X5500 processors. It is housed in a node with up to 32 GB
of memory and a DDR Infiniband connection at a speed of 2 GB/s is quoted which
means that a 2×12 connection must be used. The Infiniband switch is made
by Fujitsu. It has added functionality by supporting hardware barriers and
reduction functions, which can greatly speed up the corresponding MPI functions.
In the June 2010 issue of the TOP500 list,
[35], a performance of 110.6
Tflop/s out of 121.3 Tflop/s peak performance, or an efficiency of 91.2% on a
linear system of size 3,308,800. Presently this is the highest efficiency
measured in this way.
|