KELLY 4: Conclusion

• Parallelism is being exploited at three levels: pipelining of arithmetic, multiple FPUs, and multiple processors:

  0in[3em]

  

• The memory system parallelism is needed to sustain this performance on applications which don't fit into cache
• For peak performance we must write programs which are optimal at all three levels - but sub-optimal programs can still run at a large fraction of peak.

The moral:

High performance computing relies on achieving the maximum cost-effective advantage at every level.