Summary
Unique design point gives radically new capabilities
- Frontal attack on the distance to memory problem
- Low power solutions can reach performance levels unattainable through standard uni-processor based supercomputers.
- System-on-chip technology offers very high performance, low latency, very high bandwidth at very modest price - on par or less than cheapest PC's.
Wide range of applications will get high percentage of node peak. This architecture will deliver much higher fraction of peak on random memory access patterns than SMP clusters in the same timeframe.
Many numerically intensive computing applications require only nearest neighbor coupling with message passing and simple global functions.
Opportunity for tremendous breakthrough in price/performance for applications - 364 peak TF achievable with modest budget in 2-3 yrs
With shared risk, IBM & ASCI can make this happen.
- Research and development funding to start FY00
- BlueLight build separate contract in FY02