3 ms·
This is a great direction for a certain class of problems, but to call this a general approach for building a super computer ignores the biggest challenge of al
by noahdesu 9y ago
This is a great direction for a certain class of problems, but to call this a general approach for building a super computer ignores the biggest challenge of all: communication performance.
The leadership-class super computers have large budgets for network interconnects like Infiniband, because the hardest problems require ultra low-latency, high-throughput RDMA message passing between processes.
- xtreme 9y agoAbsolutely true. Embarrassingly parallel problems are generally not the most interesting, because you can just throw compute power at it until it is solved. Most applications running on actual supercomputers are communication bound.