 |
What is MDGARPE? |
|
MDGRAPE is the name of PC acceleration board for Molecular Dynamics calculation, which is connected through PCI interface with PC.
There are two boards, one is MDGRAPE-2 completed in 1999 and another is MDGRAPE-3 competed in 2005.
MDGRAPE-2 development
Upon hearing the word "MD-GRAPE", most of the general public would respond, "MD-what?". IBM Research Division and the Institute of Chemical Research (RIKEN) in Tokyo have been collaborating few years from 1996 to produce an accelerator chip that can rapidly calculate all of the interatomic forces in a molecular dynamics simulation with millions of particles -- a task that would take a room full of conventional computers. These chips were combined with other special chips at RIKEN to make a Molecular Dynamics Machine, which runs at a whopping 78 Teraflops speed, the fastest for any computer in the world at the time. For comparison, the fastest PC runs at about one one-hundred-thousandth of this speed. The MDGRAPE chip is the next in a series of GRAPE chips, which have won the Gordon Bell Award for Fastest Computer (given by the IEEE Computer Society) in 1995 and 1996, and for Best Price/Performance Ratio in 1999. MDM won the Gordon Bell Award for Peak Performance Category. MDGRAPE's speed results from the combination of many things.
MDGRAPE-3 development
MDGRAPE-3 was developed by Genomic Science Center in RIKEN Yokohama as a succesor of MDGRAPE-2. PCI-X board mounting two MDGRAPE-3 chips has 330 Gflops speed that is about 5 times faster than MDGRAPE-2.
MDGRAPE-3 system with 5 thousand MDGRAPE-3 chips in RIKEN Yokohama performed 1 Petaflops in June 2006.
TOP
|
|
|
|
Many physical problems can be simplified by approximating the real world with an array of particles, each representing a small piece of the total space. In the case of chemical reactions or the intricate workings of proteins and enzymes, these representative particles are the individual atoms of the molecules involved. For fluid dynamics problems, such as weather forecasting, numerical wind tunnel experiments on cars, ships and aircraft, turbulence, and so on, the model particles can be imaginary pieces of the fluid, written either as points that carry along with them the pressure, temperature, and other fluid properties, or as vectors (arrows), that mark the local vorticity (spin). Even galaxies in space, which may contain thousands of millions of stars, can be represented fairly accurately by only several million "star" particles in a simulation. For these types problems, a computer has to do the same time-consuming operation over and over again: calculate the distances between all pairs of particles, and from these distances, calculate the mutual forces between all pairs, and then sum up these forces for each particle. This calculation involves a number of operations that is equal to the square of the number of particles, and this square rapidly grows so large with large simulations that a conventional computer will just grind to a virtual halt. MDGRAPE solves this problem by specializing in just this mutual force calculation and summation, and it does this part at extreme speed.
TOP
|
|
|
|
Molecular Dynamics : it calculates any forces specified by the user, but existing libraries handle the Coulomb and van der Waals forces, and in addition to all of the real-space operations involved with the Ewald method.
Plasma Physics (charged particle interactions)
Self-gravitating systems, including cosmology, galaxies, and planets Hydrodynamics (using Smoothed Particle Hydrodynamics or the particle-vortex method)
TOP
|
|
|
|
MDGRAPE boards attach to a normal computer by fitting into the PCI slots. The host computer and the MDGRAPE board work together to solve the physical problem. The host, which can be anything from a small PC to a giant array of parallel workstations, does all of the work involving single particles, such as moving them along in time, but after each time step, the host sends the most recent particle information to the MDGRAPE boards. These boards automatically calculate and return to the host all of the forces on each particle. If desired, the boards can also return the energy of each particle. These forces are then used by the host to move the particles along, until the next time step begins.
One source of MDGRAPE's attractiveness is that it runs independently at the command of a host computer, prompted by FORTRAN or C subroutine calls that are easily embedded into standard software. You can download the subroutines from RIKEN'homepage.
TOP
|
|
Why is MDGRAPE fast? (In case of new MDGRAPE-3) |
|
Each MDGRAPE-3 chip has twenty pipelines that independently do a sequential stream of calculations following a 250MHz clock signal. Each pipeline holds 2 particle positions. Each MDGRAPE-3 chip also has a memory for particle where 32 thousands particle positions, along with their charges can be stored. All of these positions and charges are loaded onto the MDGRAPE-3 board with subroutine calls, using special library routines on the host. The list for pipeline particles can differ from the list for particle memory.
Following another subroutine call from the host, position information are written to the pipeline particle list. The pipelines automatically calculate the particle separations and then evaluate and sum the forces (or energies), fourth order polynomial whose coefficients are supplied by library functions for standard forces (van der Waals, Coulomb and Ewald-sum terms), or which may be supplied by the user through another subroutine call to represent any other force.
When this streaming is finished, the force sums return to the host and another forty particles are assigned to each pipeline. The board particles stream by again, producing new force sums for these new articles. After all of the particles on the board have streamed by all of the particles on the pipeline list, the operation is returned to the host computer.
TOP
|
|
What are MDGRAPE-3's Properties? |
|
- Project Leader : Dr. Makoto Taiji,
High-Performance Molecular Simulation Team
Genomic Science Center,RIKEN,Yokohama
- Primary Chip Design : RIKEN
- Chip Manufacturer : Hitachi
0.13 um HDL4N
10 million gates
250 MHz system clock
6Mbits RAM on chip for particle memory
165 Gflops per chip in 20 pipelines
TOP
|
|
|