Synchronization
When a memory access must wait, it is retried several times in hardware before issuing stream traps (is interrupted).
- Does not interfere with instruction issuing by other streams
- Work done by memory functional unit
- Retry count limit specified by user (retry limit register associated with each protection domain).
- Each processor has 16 protection domains (treated as separate address spaces)
Architecture supports an atomic fetch-add which can update memory locations
- Add/Sub value from a memory location atomically
- Simplifies compiler optimization