Development of the MD-GRAPE,a High-Speed Accelerator with a Parallel Processor Design for Molecular Dynamics Simulations, and its Application to Studies of Biological Membranes

由于其巨大的计算复杂性，分子动力学(MD)模拟用于研究生物大分子这样的大系统受到计算机资源可用性的严重限制。随着感兴趣系统的大小增加，要计算的非键力(库仑相互作用和范德华相互作用)的数量增加为O(N2)，其中N是系统中的粒子数量。切断库仑相互作用会造成各种有害影响，应该避免。在涉及10,000多个粒子的MD模拟中，非粘结力的计算消耗了99%以上的CPU时间。为了克服这一问题，我们开发了一台插入工作站的专用并行机，以加快非键合相互作用的计算。这台机器是一个可伸缩的同构多处理器，称为“MD引擎”。机器中的每个处理器元件都是一个由大约130,000个栅极组成的LSI芯片，采用0.8 mm CMOS标准单元工艺制造，具有流水线架构，可以使用主机广播的每个粒子的坐标、电荷和物种来计算总的非键合力。在计算力之后，处理器将其发送回主机。MD引擎还可以与压力计算中使用的力同时计算维里数，适应周期性边界条件，并可用于埃瓦尔德求和。优化了处理器内部算术运算的精度，计算出的力具有足够的精度，可用于实际的MD模拟。使用由24个处理器组成的MD引擎系统并插入SPARCStation 10/51，将浸泡在水球(11,940个颗粒)中的Ras p21蛋白质分子的MD模拟加速了75倍。

Because of its great computational complexity, the use of molecular dynamics (MD) simulations for studying large systems like those of biological macromolecules is severely limited by the availability of computer resources. As the size of the system of interest increases, the number of non-bonded forces (Coulombic and van der Waals interactions) to be calculated increases as O(N2), where N is the number of particles in the system. Cutoffs of Coulombic interactions cause various detrimental effects and should be avoided. The calculation of non-bonded forces consumes more than 99% of the CPU time in and MD simulation involving over 10,000 particles. To overcome this problem, we have developed a special-purpose parallel machine that is plugged into a workstation to accelerate the calculation of non-bonded interactions. The machine is a scalable homogeneous multiprocessor called an "MD Engine". Each processor element in the machine, an LSI chip of about 130,000 gates fabricated with 0.8mm CMOS standard-cell technology, has a pipeline architecture to calculate the total non-bonded force using the coordinates, electric charge, and species of each particle broadcast by the host computer. After the force is calculated, the processor sends it back to the host. The MD Engine also calculates virials simultaneously with forces for use in the calculation of pressure, accommodates periodic boundary conditions, and can be used in Ewald summations. The precision of arithmetic operations inside the processor is optimized, and the force is calculated with sufficient accuracy for practical MD simulations. MD simulation of a Ras p21 protein molecule immersed in a water sphere (11,940 particles) was accelerated by a factor of 75 using an MD Engine system consisting of 24 processors and plugged into a SPARCstation 10/51.