As the ASC Student Supercomputer Challenge (ASC16) enters competition, on January 26, ASC announced that the “e Prize ” this year would set on a deep learning algorithm DNN (Deep Neural Network). All teams are to develop DNN on a MIC (Many-Integrated-Core) platform and optimize the application on a provided MIC supercomputer cluster.
DNN is one of the most important deep learning algorithms at present. It aims to construct a set of autonomous learning neural network that simulates the human brain, and thus to realize deep artificial intelligence. Many companies like Google, Microsoft, Baidu, and iFLYTEK have utilized DNN algorithm and a heterogeneous acceleration technology in their artificial intelligence research. However, the mostly used DNN algorithm is GPU heterogeneous architecture. ASC16 requires the teams to move DNN on MIC. The students won’t have extant successful cases for reference. They will have to rely on their own innovative thinking to address this challenge.
“e Prize ” is a major award exclusively provided by the ASC. “e” represents not only the most important natural constant in science, but also the next challenging target for the supercomputer performance — exascale. ASC hopes that the e Prize can become the Gordon Bell Prize for youth talents, with the aim of encouraging them to be brave in overcoming the difficulties on supercomputer application and challenging the limit of computing performance, motivating them to use the supercomputer to facilitate the realization of application breakthroughs in the fields of science and engineering. Gordon Bell Prize is jointly awarded by Association for Computing Machinery (ACM) and Institute of Electrical and Electronics Engineers (IEEE) for commending the achievements made by the researchers in making scientific breakthroughs via parallel computing. This prize began in 1987 and ranks among the top three in the U.S. engineering field with the reputation of “Nobel Prize in the Supercomputer Field”, the other two being Charles Stark Draper Prize and Russ Prize.
In fact, “e Prize ” gives more weight to “Challenge” which is all about improving the concurrent design and efficiency of scalability of supercomputer application software to truly give full play to the application value of the supercomputer and change the situation of “Great machines being used for mini-applications” universally confronting the development of supercomputing.
The previous competitions suggest that “e Prize” can really motivate the zeal of college students for innovation and has achieved good results in facilitating the development of large-scale supercomputing applications. In ASC13, the team of Huazhong University of Science and Technology achieved an acceleration effect of 60,000 times on BSDE (Backward Stochastic Differential Equation). Among the results of speed-up ratio of 60,000 times, more than 300 times resulted from such optimizations as parallelization, vectorization, memory access, etc. and more than 200 times from the optimization of algorithms, which fully demonstrated the innovative capabilities of the college students and their deep understandings to the applications. BSDE has been extensively studied in the field of financial engineering in recent years. It can not only be used for the more precise and more practical pricing calculation and analysis, but also help various types of investors to perform the risk hedging and other risk analyses.
The test of the “e Prize ” in ASC14 was about the 3D-EW (a simulation method for P-and S-wave separation-based 3D elastic wave equations) provided by Bureau of Geophysical Prospecting INC., China National Petroleum Corporation. This 3D-EW features small dispersion, high computing accuracy but much computation, and has a high scalable potential. Eventually, Shanghai Jiaotong University successfully extended the 3D-EW from mononuclear to 200,000 nuclear on Tianhe-2, the fastest supercomputer in the world, and won the first prize by spending 42 minutes in the computing.
ASC15 realized the connection between “e Prize ” and the international large science project for the first time by taking Gridding, the data processing software of SKA (acronym for Square Kilometer Array, which has so far been the world’s largest international scientific research cooperation project in the field of astronomy), in the hope of gathering the wisdom of worldwide young supercomputing talents to speed up the research in exploring the universe. At last, the team of Sun Yat-sen University successfully optimized and correctly operated the all three examples of Gridding, realized an acceleration effect of 600 times, reduced the system power consumption of the data processing in the same scale to 1/20 of the original and provided an excellent advice in designing the high performance supercomputer for SKA, thus winning the first prize in the competition.
In the future, with more practical problems selected for the “e Prize “, young talent around the world are free to apply their wisdom in supercomputing to benefit the mankind.