"... In PAGE 8: ... Table4 : Practical Results As we can see from the table, SBS-0.75 is giving the best overall performance.... ..."

"... In PAGE 5: ... is similar. Table2 lists the energy consumption for a sin- gle access in each of the prediction tables, as well as for an access in a traditional LVP. The tables are sized for a hardware budget of 8KB.... In PAGE 5: ... The tables are sized for a hardware budget of 8KB. We measured the number of each kind of predictor lookup and, with the energy costs of Table2 , computed the overall energy consumption of the value predictors. Figure 6 shows the total value predictor energy consump- tion for the 8KB predictors.... ..."

"... In PAGE 15: ... Decentralized last value speculation was applied to only integer instructions in the benchmarks. Table2 lists the performance of the last value predic- tor across the set of benchmarks. The first column shows the IPC of the benchmark on the base case without value prediction.... In PAGE 15: ... The second column shows the speedup ob- tained when using only the 2-bit counter. We see from Table2 that using the only the 2-bit counter actually hurts the performance of some benchmarks. However, some benchmarks like adpcm and mcf, show apprecia- ble speedup with the last value predictor.... In PAGE 15: ... We found that the low accuracy of the 2-bit counter generates a large number of mis-speculated values in the GPA resulting in ALUs firing multiple times to generate the right value. The third column in Table2 lists the speedup ob- tained with the 2-bit counter and poison bit. We see from the table that using a poison bit never hurts performance.... In PAGE 16: ...9% 5.9% Table2 . Last Value Speculation Performance on the GPA load-prediction strategies, including dependence predic- tion, address prediction, value prediction, and memory renaming [3].... In PAGE 23: ... We chose them because the selected loops exhibit considerably different behaviors with respect to the prediction patterns. Table2 shows the results of the selected loops. For each application, we have three columns: variable, pattern and ratio.... In PAGE 24: ...9 98.4 Table2 : Selective value profiling experiments results compress95 vpr twolf Input data set input.ref input.... In PAGE 31: ...3. Prediction coverage Table2 provides the percentages of static methods (Static) as well as dynamic return values (Dynamic) that are correctly predicted by the parameter stride pattern. In addition, the table shows the percentage of overall return values that are correctly predicted by the PS prediction but not by all other predictors (Uncovered).... In PAGE 31: ... On average, about 13% of dynamically encountered return values exhibit a predictable parameter stride pattern, and 47% of these return values can not be predicted by other types of return value predictors. Table2 . Percentage of Methods and Return Values that Possess the Parameter Stride Pattern (sequential execution) comp db javac jess mpeg mtrt average Static 13.... In PAGE 44: ...3% 76.1% Table2 : Benchmark Statistics Table 2 lists the statistics of the benchmarks, on the number of L2 cache misses, shared load misses, and shared load misses by non-atomic accesses. It also shows that 27% to 45% of L2 cache misses are R-misses, and 70% to 80% of the shared load misses are made by nor- mal non-atomic load instructions.... In PAGE 44: ...3% 76.1% Table 2: Benchmark Statistics Table2 lists the statistics of the benchmarks, on the number of L2 cache misses, shared load misses, and shared load misses by non-atomic accesses. It also shows that 27% to 45% of L2 cache misses are R-misses, and 70% to 80% of the shared load misses are made by nor- mal non-atomic load instructions.... In PAGE 50: ... The genetic algorithm can therefore be used to verify manual design choices. Table2 shows the misprediction rates (lower is better) of manual and evolved (400 generations) two- level predictors employing an ideal table, a 1K-entry fully associative table and a 1K-entry tagless table. An ideal table is a fully-associative table of infinite size.... In PAGE 50: ...6.3 15.2 15.5 Table2 . Misprediction rates for manually designed and evolved two-level predictors.... In PAGE 50: ... Misprediction rates for manually designed and evolved two-level predictors. The last two columns of Table2 show the misprediction rates for evolved two-level predictors with a maximum path length of 6 and 24. Convergence to a near-optimal solution is much faster for the former, since bits from the 7th target or later rarely improve prediction accuracy.... In PAGE 75: ... The parameters for the simulated base microar- chitecture are in Table 2. Table2 : Baseline Microarchitecture Configuration Datamemory L1DataCache: 4-way,32KB,2-cycleaccess L2Unifiedcache: 4-way,1MB,10cycle Non-blocking 12MSHRsand2ports D-TLB 512-entry,4-way 1-cyclehit,30-cyclemiss Storebuffer: 32-entryw/loadforwarding Loadqueue: 32-entry,nospec. disambiguation MainMemory Infinite,75cycle FetchEngine Tracecache: 4-way,1Kentry,3-cycleaccess 16instr.... In PAGE 90: ... is similar. Table2 lists the energy consumption for a sin- gle access in each of the prediction tables, as well as for an access in a traditional LVP. The tables are sized for a hardware budget of 8KB.... In PAGE 90: ... The tables are sized for a hardware budget of 8KB. We measured the number of each kind of predictor lookup and, with the energy costs of Table2 , computed the overall energy consumption of the value predictors. Figure 6 shows the total value predictor energy consump- tion for the 8KB predictors.... ..."

"... In PAGE 7: ... 4.7 Experimental Results The experimental results are shown in Table5 . The exe- cution time Tr and energy consumption Er are all relative to the case in which the same program was run on the non-DVS system, i.... ..."

"... In PAGE 7: ... 4.7 Experimental Results The experimental results are shown in Table5 . The exe- cution time Tr and energy consumption Er are all relative to the case in which the same program was run on the non-DVS system, i.... ..."

"... In PAGE 7: ... 4.7 Experimental Results The experimental results are shown in Table5 . The exe- cution time Tr and energy consumption Er are all relative to the case in which the same program was run on the non-DVS system, i.... ..."

"... In PAGE 12: ... This means that FLIP-optimized dif- fusion could double the lifetime of a sensor network when compared to plain diffusion. Table3 summarizes en- ergy consumption results for the different diffusion vari- ants. Table 3: Diffusion variant energy consumption... ..."

"... In PAGE 12: ... This means that FLIP-optimized dif- fusion could double the lifetime of a sensor network when compared to plain diffusion. Table3 summarizes en- ergy consumption results for the different diffusion vari- ants. Table 3: Diffusion variant energy consumption... ..."