[Updated: 6/13/2011 - Changed 25 microseconds to 25 milliseconds]
When the National Supercomputing Center in Tianjin, China, announced that they built Tianhe-1A — the world’s fastest supercomputer based on NVIDIA Tesla GPUs — there were a lot of folks who questioned whether this system was just a stunt machine or a true deep investment in science and technology.
Well, it looks like GPU supercomputers, like the Tianhe-1A, are just a small part of the investments that China is making in science and technology sectors – especially the high-performance computing (HPC) space. Tianhe-1A has proven today that it’s more than just a record-breaking machine – it’s a true scientific tool.
Scientists at the Institute of Process Engineering (IPE) at the Chinese Academy of Sciences (CAS) today announced a record-breaking scientific simulation on the Tianhe-1A GPU supercomputer that furthers their research in solar energy. CAS-IPE scientists ran a complex molecular dynamics (MD) simulation on all 7,168 NVIDIA Tesla GPUs to achieve a performance of 1.87 petaflops per second – that’s about the same performance as 130,000 laptops. These scientists are simulating the structure of crystalline silicon which is used in solar panels and also in the semiconductor industry. And they were able to accomplish this world’s fastest MD simulation by writing just 2,000 lines of CUDA code.
The Tianhe-1A supercomputer. Powered by NVIDIA Tesla GPUs.
Some technical facts for the scientists among you (everyone else can safely skip this paragraph):
- The researchers used a multi-body, bond-order potential (BOP) application that they developed at CAS-IPE.
- The simulation was run using single precision floating point calculation since it was sufficient for their problem. CAS-IPE researchers verified this by calculating various physical quantities from the end result and found them out to be physically correct against actual materials.
- They simulated a system with 110.1 billion atoms and the simulation ran for about 500,000 time steps to simulate a physical time of 0.116 nanoseconds of the system evolution (it seems this is a meaningful time span for the physical problem at hand). Each step takes 25 milliseconds to compute, so the whole simulation takes about 3 hours on the Tianhe-1A system.
China wants to expand their economy from being manufacturing-based to also focusing on product design. Supercomputing enables better science and engineering, which in turn leads to higher quality product design along with faster manufacturing, and reduced costs. Policy makers in China realize that supercomputing is a fundamental tool in the global competition for economic growth, and are pursuing GPU-based supercomputers to realize their goals.
This record-breaking simulation of the materials used in solar panels is just a first step. We’ll keep you posted on the next giant leaps in GPU supercomputing.
Tagged: CUDA, parallel computing, Supercomputing, Tesla
