diff options
Diffstat (limited to 'content/zettel/1c6e.md')
| -rw-r--r-- | content/zettel/1c6e.md | 24 |
1 files changed, 24 insertions, 0 deletions
diff --git a/content/zettel/1c6e.md b/content/zettel/1c6e.md new file mode 100644 index 0000000..ac07ade --- /dev/null +++ b/content/zettel/1c6e.md @@ -0,0 +1,24 @@ ++++ +title = "Speculation does not have a cost" +date = "2022-04-29" +author = "Yann Herklotz" +tags = [] +categories = [] +backlinks = ["1c6d"] +forwardlinks = [] +zettelid = "1c6e" ++++ + +It is interesting to note that speculating on an instruction does not +really incur a cost in itself. The idea is that if one is always wrong +with the speculation, one really only reverts back to the non-pipelined +instructions. So in the worst case one is only sequential. However, if +one speculates, then it is highly unlikely that one is always wrong, and +one will therefore always benefit from it. + +Therefore, in the speculative loop pipelining, it does not really matter +in the end which instructions are picked to speculate on, and often the +most efficient will be the shortest path in the pipeline. For example, +Steven showed that even when one speculates on a binary search algorithm +to two iterations (where one then has a 25% chance of being right), one +still benefits from speculating going into one of the directions. |