Building a Petaflop computer from scratch is not that difficult

Speaking in terms of power, the concepts of giga, tera, peta, exa and their counterparts, like nano, pico, femto, and atto, are closely related from a technological as well as investor's perspective. If we are to design a Petaflop system from zero, we have to take into account a series of important basic rules, like the fundamental laws of physics, and the developing trends in telecommunications, semi-conductors, and computer architecture. The interaction between those factors eventually will lead us to Petaflop computing. As far as the physical laws are concerned, elements such as the speed of light, the overall power consumption, and the propagation delay enter into play.

As Mr. Wallach explains, there is a well defined limit to raising power by increasing the frequency which makes the engineer stop before the circuits are melting down. In the same way, we face built-in constraints with regard to the propagation delay. The cost of investment and the market size form yet another issue to be solved. If we divide the millions of market resources through the billions of investment dollars, we arrive at a result which is whether targeted towards success or hampered by failure. This is what he refers to as L'Hospitals rule of profit.

The indispensable ingredients for Teraflop system construction include photonic technology, in particular Wavelength Division Multiplexing (WDM), TeraHz requirements, and the emergence of All Optical Networks (AON), together with their impact on digital computer architecture. As Mr. Wallach states, we have to examine what is driving the investment in the equation for success. The costs for the "FABS" or production units escalate to over $2 billion. Should we have one FAB per continent or do we have to build it on the moon? As a matter of fact, billions of dollars are required to replace the 8 inch fabs. However, there is some good news too since the costs of the chips will be reduced.

Mr. Wallach holds in store a well furnished scenarios to succeed in establishing the a priori set out goals which turns the message very clear: no promises for a rose garden. The thing is how we intend to use the PC/personal workstation of the future, consisting of four elements, namely the risc core, the memory unit, referred to as DRam/Cache, the graphical unit, and the parallel functional units. Choices will need to be made in favour of the parallel functional units or the graphical unit.

Talking about high performance computing meabs touching the Linux issue. By 2001, the Linux basic code will be as big and and complex as Unix is today and, because of its exponential growth rate, will approximate Windows 2000's size in nuber of lines of code in another few years. Mr. Wallach predicts a Processor in Memory (PIM) or a System-on-a-chip (SOAC) architecture, with more memory bandwidth and a lower latency, and showing some consistency with the PC pricing and technology curves. From the software point of view, interconnection will be enabled through message passing, Distributed Shared Memory (DSM), Cache Only (COMA) or object oriented technologies, or emulated DSm, referred to as threadmarks.

As far as hardware is concerned, the future lies in a physical combination of copper and photonic. Eventually, WDM will take the lead in external chip interconnects. In a 2 to 4 years short term vision, the accent will be put on the SOAC architecture whereas in the long run, we will evolve towards a Petaflop system. The core unit will consist of an All-Optical data or communication switch that is interconnected with a number of Shared Memory Processors. The physical memory will include a cache, a local as well as a global memory. In addition, there will be an input/output unit and the provision of a substantialm sustainable bandwidth, just like in vector processing.