Our approach to high-performance computing is to use a special-purpose computer for a specific problem. Our current targets include molecular dynamics (MD) simulation and homology search. We have accelerated the MD simulation by developing the world's fastest machine, called the Molecular Dynamics Machine (MDM). We have also accelerated the homology search by using a field-programmable gate array (FPGA). Our research on MD simulation is described in more detail below.

The MDM is a special-purpose computer for large-scale molecular dynamics (MD) simulation. It is composed of two special-purpose computers, MDGRAPE-2 and WINE-2, and a host computer. It has a peak speed of 78 Tflops, which was six times as fast as the world's fastest supercomputer in 2001. We received the Gordon Bell prize in the peak performance category in 2000 for our MD simulation with the MDM. In 2001, we achieved a 35-Tflop MD simulation with 67 million atoms. We believe this is the fastest calculation ever done in a molecular dynamics simulation. The MDM was developed at the Computational Science Division, Advanced Computing Center, RIKEN. and papers on MDM are available through the Internet.

We are now concentrating on the following issues.

• Protein simulation with the MDM

We are simulating proteins, such as Ras p21 and prions. MDM can calculate the Coulomb force without a cutoff, so we can do very accurate simulations even of very large systems with more than 100,000 atoms. You can see movies of our simulations through the Internet.

• Porting of MD software to the MDM

We have ported AMBER and CHARMM, which are widely used MD software packages for protein simulation. Researchers can now use the MDM without being concerned with the hardware. Some functions of these software packages have not yet been implemented for the MDM, so the porting is continuing. You can see the option 'GRAPe' in the 'method-spec' of the non-bond options of CHARMM (at about the 70th line).

• Development of new hardware

  We plan to build a new machine, which will reach one or two petaflops, within the next several years. We are also modifying the existing hardware to enhance its performance.