(5th-February-2020)
Figure 2 Example of high-precision simulation using molecular dynamics (MD) calculation
The figure shows an example of a high-precision simulation of a protein obtained by MD calculation. The molecule represented by the ribbon model represents a protein, and the molecule represented by the ball and stick model represents a ligand. When calculating the ease of binding by calculation, since both proteins and drug molecules actually change shape, it is important to capture the change. In MD calculations, proteins, molecules serving as drugs, and surrounding water molecules are modeled at the atomic level, and their time evolution is followed.
Research methods and results
The research team built an accelerator specializing in calculating the force between particles required for MD calculation into a large-scale integrated circuit (LSI), and developed the MD simulation supercomputer "MDGRAPE-4A" equipped with 512 LSIs. (Figure 3). MDGRAPE-4A has the ability to simulate a 100,000 atom system consisting of protein and water molecules for up to 1.1 microseconds per day in a calculation. As a result, the computation time required for a movement of 100 microseconds is 91 days, and a simulation that took at least one year and three months with a general-purpose supercomputer can be completed in about three months.
Until now, the dedicated computer developed by RIKEN had dedicated only a part of the calculations, and the rest was calculated using ordinary computers. However, as the speed of dedicated computers has increased, the performance of this method has reached a plateau in ordinary computers. Therefore, in MDGRAPE-4A, in addition to the conventional dedicated calculation circuit, a large-scale “system-on-chip (SoC) [9]” that integrates all calculations such as general-purpose calculation parts and networks into one LSI, We are working to eliminate bottlenecks. This required a lot of new technology development. The main ones are:
• (1) Development and hardware implementation of calculation algorithms suitable for dedicated computers to accelerate calculation of forces acting between distant atoms.
• (2) A high-speed dedicated circuit that calculates the force acting between nearby atoms.
• (3) A high-speed, low-latency network that connects 512 LSIs with optical fibers.
• (4) RISC-V [10] processor modified for MD calculation.
• (5) Memory in which arithmetic units and data management circuits are embedded.
• (6) Implementation of ultra-high-speed three-dimensional FFT [12] using FPGA (programmable integrated circuit) [11].
In addition to these individual ideas, we worked on the co-design of hardware and software so that all of the elements implemented in the LSI could operate at high speed. Furthermore, by assembling 512 LSIs as a system, the system as a whole has a calculation capability of about 1.3 petaflops (1,300 trillion operations per second), and has achieved a system operation that can perform high-speed calculations. This system is the world's first practical large-scale system based on RISC-V.
Comments