Information Technology Reference
In-Depth Information
Step 0
Step 0
WAY
Predictor
WAY
Predictor
DATA
DATA
DATA
DATA
DATA
DATA
DATA
DATA
Step 1
Step 1
Step 2
Compare
Step 2
Compare
MUX
MUX
Step 3
Step 3
DATA
DATA
DATA
DATA
DATA
DATA
DATA
DATA
Step 1
Step 1
Step 2
Compare
Step 2
Compare
MUX
MUX
Step 3
Step 3
DATA
DATA
DATA
DATA
Step 1
Step 2
Compare
MUX
Step 3
FIGURE 4.27:
Hybrid, way prediction
phased access. On the left, way prediction is combined with
phased access. The phased access takes place only on mispredictions. The remaining tag ways are accessed
in the misprediction tag phase and the correct data way is accessed in the misprediction data phase. On
the right, the way prediction and the misprediction tag phase are combined into one. All tag ways are
accessed in the first phase along with the predicted data way. In the case of a misprediction, the correct
way is accessed in the misprediction phase.
+
according to where the data are found on hits or with the location of the replaced data on misses.
Inhue et al. provide evidence that MRU is sufficiently good prediction in 4-way associative
caches to outdo a phased cache in EDP for various SPEC2000 benchmarks.
The power reduction of this scheme is proportional to the prediction accuracy and
inversely proportional to the associativity:
P
new
=
Accuracy
×
P
old
/
Ways
+
(1
−
Accuracy)
×
P
old
.
0) this scheme cannot be worse in power consumption than a
standard cache, but it certainly will be slower (i.e., worse EDP than the base case). Unfortunately,
the two factors that determine the power reduction are antagonistic. Higher associativity yields
larger benefit on correct predictions, but at the same time lowers the MRU prediction accuracy
(resulting in fewer correct predictions).
An improvement in this scheme is to combine it with the phased access approach
(Figure 4.27, left diagram). The MRU tag is checked first and in case of a hit the corresponding
At worse (Accuracy
=
Search WWH ::
Custom Search