Page History

...

1. feedback-first: 
   +it is simple and most predictable 
   +discourages spilling
   +lowers latency per iterative operation
   - lowers throughput -> stages loops and discourages pipelining
   - restricts loop logic to staged execution (waiting until all cyclic operations are done and output is flushed out of the loop)
   - it can create some form of temporary starvation when the feedback queue is constantly non-empty. This can stale fault tolerance barriers and other events that serve operational system needs.
2. feedback-first when feedback queue buffers are utilised:
   
   + discourages spilling
   + lower end-to-end latency and higher throughput
   +/- it limits the possibility of temp starvation only when feedback queue is excessively morebusier
   - it can increase latency per-loop operation due to multiplexing
   - less predictable and more complex conceptually

...