Conclusions
The SV1 offers very good performance for vector applications
Still very much a vector machine
New optimization issues naturally center on
As expected, code performance closely tracks memory throughput
Future PE releases should enhance performance of the MSPs
- Improved loop restructuring
- Reduction of synchronization overhead
- Multi-streaming of constructs in addition to loops
- Additional directives, etc., to give programmers more flexibility in optimization
We eagerly await these developments, not to mention the SV2!