A developer of next-generation super computers
"I aim to deliver supercomputers that provide real value to customers"
NEC is the only development vendor in the world continuing development of vector supercomputers. Why the focus on vector-based technology? What is the new "SX-ACE" supercomputer? How can supercomputers contribute to society in the coming years? Shintaro Momose, a developer of next-generation supercomputers, talks about the true value of supercomputers, as well as the appeal of state-of-the-art machines. He goes on to discuss fields in which supercomputers show promise, as well as what he finds rewarding as an engineer.
Supercomputers are seeing increasing use at companies
--First, tell us about recent trends in the field of supercomputers, and how they are being used.
Momose: Until about 10 years ago, supercomputers were mostly used for scientific computations and the ultra-high speed simulation of physical phenomena in cutting-edge research fields at universities and research institutes.
Later, the appearance of versatile processors such as the Xeon developed by Intel led to the rise of cheap supercomputers involving massive banks of PC-like servers called PC clusters. Because of this, the use of supercomputers has spread like wildfire to areas other than cutting-edge research, such as car and aircraft manufacturers, pharmaceutical companies, major e-commerce sites, and film production companies.
Recent trends have seen a split into two categories. Performance-oriented supercomputers developed for special purposes are used in cutting-edge research at universities and research institutions. Meanwhile, more cost-oriented PC clusters are seeing widespread use at companies and other organizations.
--What trends are you seeing among supercomputer development vendors?
Momose: The history of supercomputers began with vector-based machines adept at large-scale computation, which were developed by U.S. company Cray. Many vendors both in Japan and overseas also once developed vector-based systems. However, with the advent of versatile scalar-based processors, most vendors switched to developing scalar supercomputers.
NEC is now the only vendor in the world still developing vector supercomputers. The competition for the top supercomputer often draws attention, but I don't believe the value of a supercomputer can be determined simply based on which is the fastest. For example, an athlete who can run a fast 100 meters will not necessarily also be good at playing soccer or baseball. The same goes for supercomputers. Speed is certainly also important, but I think the true value of a supercomputer lies in achieving the best performance that meets a range of objectives.
There has previously been debate surrounding whether a scalar or vector approach was best, but there is little point to this these days in my opinion. That is because vector and scalar approaches each have different strengths. Vector systems are adept at the high-speed processing of large-scale datasets, while scalar systems are more suited to processing smaller datasets in large numbers. There are a variety of reasons for using supercomputers, as well as the characteristics of the applications they run, so I think the most crucial thing is to use the right product for the right application.
Vector systems excel at the high-speed processing of large datasets
--Tell us about the history and current status of NEC's supercomputer development.
Momose: NEC's supercomputers consist of the vector-based SX series, but we also develop and offer scalar-based products that use general purpose processors for the HPC (High Performance Computing) area. Our vector-based SX series supercomputer line was launched in 1983, and the latest state-of-the-art SX-ACE model is a tenth generation product. One defining feature of NEC's approach is that we develop all the components that make up our supercomputers ourselves, from the CPU, to the hardware, OS, and software.
--Can you explain why NEC has stuck with vector-based systems?
Momose: While scalar systems have the upper hand in catalog performance, when it comes to the ability to actually run applications, vector systems perform very well. In recent years there has been new demand for the high-speed processing of large-scale data, including big data, as well as applications in specialist areas such as fluid dynamics and meteorology.
These days, vector functions are added to scalar supercomputers to improve performance in most cases. That means that improving vector performance is the most important way to achieve high effective performance, even with scalar processors. In that respect, it could be said that vector processors are ahead of the trend.
I think vector-based systems that excel at the high-speed processing of large datasets, such as in fluids dynamics or meteorology, will also demonstrate their ability in the realm of big data. There are high expectations for their application in this area. NEC's major advantage over other development vendors is the accomplishments and know-how we have built up over more than 30 years of developing vector supercomputers from scratch, including hardware and software. The technical expertise we have accumulated through years of developing the world's fastest processor cores can't be overstated.
NEC supercomputers forecast the weather in Europe
--Can you give us some examples of how NEC's SX series supercomputers have been applied?
Momose: The SX-9, which is a generation behind our latest model, was used in the Earth Simulator deployed at JAMSTEC, as well as at other institutions such as Tohoku University and Osaka University. Tohoku University also employed the SX-9 in simulations for the development of the MRJ, Japan's first passenger jet.
Overseas, the SX-9 has been deployed at Germany's University of Stuttgart for academic use, as well as at meteorological services in Germany and France. Our customers in Europe have come to rely on the SX series, and I've heard some say they consider it an essential tool for their weather forecasts.
--What issues have there been with vector supercomputers in the past?
Momose: Vector supercomputers have a reputation for being costly and consuming too much power. NEC places the utmost importance on our design concept for the SX series, aiming to make it easy for customers to use them as tools for their research, and not just pursuing high performance single-mindedly.
Focusing on ease-of-use to alleviate the burden of programming or code development for customers inevitably leads to higher costs. That's why it is up to the customer to decide whether to prioritize cost or ease-of-use.
Regarding power consumption, scalar solutions win out when just looking at the catalog specs, but for the areas that the SX series is targeted at, it boasts top-class energy efficiency when actually running applications.
NEC meets a range of HPC needs
--Tell us about NEC's initiatives toward the HPC business, including supercomputer development.
Momose: HPC is a moniker traditionally used for supercomputers focused on scientific computation, but recently HPC use has been spreading to new areas such as big data applications and high-speed transaction processing at companies.
NEC currently provides SX series vector supercomputers suitable for large-scale data processing, as well as scalar HPC-oriented products that incorporate general purpose processors. We are also able to offer hybrid systems that combine vector and scalar products.
In other words, with the wide range of products NEC has available, we may be the only manufacturer able to flexibly accommodate all our customers' needs. The new SX-ACE model we developed will play a key role in NEC's HPC business.
--Can you fill us in on the specifics of what your job entails with regard to the HPC business?
Momose: It is said that supercomputers achieve a thousand-fold increase in speed every 10 years. I am involved in development of NEC's flagship SX series server products, and my task is to create products while keeping an eye on the future four or five years ahead.
In five years' time we will need to aim for 30 times the processing power of today's products. The development of supercomputers incurs tremendous costs and requires lengthy development cycles. We must consider how technology will evolve in the next five years, what prices and power consumption to aim for, how other development vendors will fare, and how the environment and trends will change for users and society. My most important job is identifying a range of requirements based on these forecasts, and performing exhaustive investigations so NEC can create supercomputers that represent true value five years down the line.