Design Twist for Supercomputers: Flash Memory
It’s a heady time in high-performance computing, driven by rapid improvements in the chips that supply number-crunching power. The next target could be data storage.
For the past 15 years or so, most of the action has been in packing hundreds or thousands of PC-style microprocessors into supercomputers called clusters. More recently, some researchers have been trying to boost performance further by adding graphics chips, which each have hundreds of specialized processors. A system from China, for example, this month took the No. 2 spot on a semiannual ranking of the 500 largest systems–historically dominated by supercomputers in North America–by using a combination of Intel (INTC) microprocessors and Nvidia (NVDA) graphics chips.
But it doesn’t make sense to have ultra-speedy calculating engines if they have to wait around for the equivalent of fuel–the data they need solve problems. So engineers also keep trying to design faster connections between microprocessors, and to connect them with chips called DRAMs (dynamic random access memories) that serve as a temporary scratch pad for data and the disk drives that provides longer-term storage for information.