Results

As a result of a detailed examination of the issues, technology projections, and alternative architecture structures, the working group produced key findings that should establish the direction for future research leading ultimately to effective PetaFLOPS-scale computation. The major findings were:

1. A PetaFLOPS computer architecture will be feasible in a 20 year time-frame.

2. No new architecture paradigm is required to achieve PetaFLOPS computing. Highly advanced versions of today's multiprocessor architectures, combined with known techniques, should provide the basis for PetaFLOPS computing systems.

3. Memory will dominate in determining system size, structure, and cost. An important finding is that scientific problems employing a PetaFLOPS system will not, in most cases, require a petabyte of memory, but between one and two orders-of-magnitude less, thus significantly reducing the potential size and cost of the system.

4. Memory latency and bandwidth will be the most critical factors determining effective performance and will require a radical departure from today's typical methods of processor-memory interaction.

5. The driving factor determining the rate of performance evolution is market forces. These forces, resulting from mass market computing requirements rather than those of high-performance computing, will determine when the necessary components will be available and their functionality.

6. Although semiconductor technology still will be the source of the majority of components, cryogenic superconducting technology may provide the processing rates required for the first and possibly the second of the architecture types considered. Also, an important consideration is power consumption, which is extremely low for superconducting devices. Optical devices are unlikely to provide high-speed logic but will provide the bulk of the interprocessor and memory bandwidth. A concern is that SIA projections do not extend far enough and extrapolations upon which these conclusions are based are subject to revision.

7. Reliability will be determined largely by parts count and this will be limited by cost factors. System costs will constrain parts (i.e., chips) count to a range of between a hundred thousand and a million. This value is at the high end of today's largest systems but should be manageable through advanced engineering techniques.