Case Study:
NTU: Chess-playing Supercomputer

It takes the world's fastest computers to beat the world's greatest chess players. Before the IBM Deep Blue finally beat the Russian Grand Master, Gary Kasparov, in a highly publicized second encounter, there had been a great rush to build the ultimate chess-playing supercomputer. The real life applications of massively parallel processing had prompted the world's premier universities to spend resources in building chess-playing supercomputers, since that epitomized the development of computing technology.

While Feng-hsiung Hsu, Murray Campbell and Thomas Anantharaman were building Deep Thought and then Deep Blue at Carnegie Mellon University, a team from Singapore's Nanyang Technological University department of Electrical Engineering decided to make their own chess-playing machine. The NTU team had considered using 128 Transputer processors designed and manufactured by Inmos company (later acquired by SGS-Thomson). Each Transputer would have had 4 link ports to communicate with other Transputers, with each link port limited to 20 Mega bit per second (20 Mb/s). Although in theory a large network of Transputers could have been used for parallel processing, in practice this would have been hindered by two bottlenecks in the form of limited data bandwidth and excessive power dissipation.

The second possibility the NTU team investigated also had similar problems. Texas Instrument's TMS320-C40 DSP with 6 parallel links was similar but more powerful than the Transputer. Although the link port bandwidth was 8 times larger then the Transputer links, it was not large enough for data intensive applications and the even larger dissipation of power would have limited the size of the network.

Understanding special needs

The difference Neurotech made to the project was in understanding the special requirements of implementing massively parallel processing for playing chess. By applying a system analysis to the Transputer and TMS320-C40 DSP proposals, we were able to advise the NTU team of the bottlenecks and limitations associated with those systems. To provide a higher performance alternative, we based our solution on a MasPar machine with an established software platform and proven scalability in processing power. Since a computer is not 'psychological', its main advantage over a human being is in its speed of evaluating chess positions. While a Grandmaster like Garry Kasparov can examine approximately three positions per second, a massivelly parallel processing machine can evaluate over 100,000,000 positions in a second. Understanding the importance of extending this advantage, we proposed that the entry level system be the MP-1101 with 1024 processors scalable to 2 thousand, 4 thousand, 8 thousand and 16 thousand processors.

Another need we addressed was in providing an easily re-configurable machine. At the time NTU got involved, the ACM competition was so competitive that some participating teams brought entirely different designs from year to year. The NTU team took confidence since the MasPar machine we provided had a balanced architecture consisting of Processor Array, Array control unit, high speed I/O subsystems and a Unix front end. With a global router bandwidth of 80 to 1,300 Mbtyes/s and a 2D Torus Grid with 8-way nearest neighbourhood communication bandwidth of 1,400 to 22,400 Mbtyes/s, our solution provided a much more flexible processing capability than the Transputer or TMS320-C40 DSP alternatives.

Although all ACM participants at that time desired flexible paltforms so as to allow them to support modified algorithms, this was even more so for the NTU team since they were still in early development stages. The direct memory bandwidth range of 680 to 22, 400 Mbytes/s and indirect memory bandwidth range of 230 to 3, 700 Mbytes/s espcially suited the NTU team, as it allowed them to build larger and more advanced programs over time. The quick implementation of our solution also helped NTU breeze through the early development phase. With installation and training at their designated site with the help of massively parallel processing experts brought over from California, we delivered the NTU chess-playing machine within 3 months.