4. EXPERIMENT DESCRIPTION AND RESULTS
4.3 Study on Intel Core i7-3970X Sandy Bridge-E Microprocessor
Benchmark
Configuration
Core 0
Core 1
Core 2
Core 3
Core Average
by Benchma
rks
All Cores
All Cores No HT
Bzip 20% 20% 20% 20% 20% 14% 14%
Mcf 0% 0% 20% 0% 5% - 17%
Milc 20% 20% 20% 20% 20% 17% 17%
Gobmk 20% 20% 20% 20% 20% 12% 14%
Zeusmp 17% 20% 17% 17% 17.75% 11% 22%
Namd 20% 20% 20% 20% 20% 12% 14%
Hmmer 20% 20% 20% 20% 20% 12% 14%
h264ref 20% 20% 20% 20% 20% 12% 14%
Benchmark Average by Cores
17.1
%
17.5
%
19.6
%
17.1
%
Table 20: Highest Power Consumption gains for all Benchmarks on all Configurations for i5-4200U
Active Cores
Configuration Core 0 Core 1 Core 2 Core 3 Core 4 Core 5
Core 0 X
Core 1 X
Core 2 X
Core 3 X
Core 4 X
Core 5 X
All Cores No HT X X X X X X
Table 21: Presentation of Configurations for i7-3970X
Benchmark
Configuration
Core 0 Core 1 Core 2 Core 3 Core 4 Core 5
All Cores No HT
Bzip 388 388 388 388 387 386 477
Mcf 213 211 212 211 212 209 352
Milc 397 395 393 395 399 399 615
Gobmk 366 365 366 374 365 365 415
Zeusmp 196 196 196 196 196 196 264
Namd 359 359 359 357 361 360 397
Hmmer 489 490 489 490 490 491 543
h264ref 449 441 444 449 449 447 517
Table 22: Timings in seconds for all Benchmarks for all Configurations on i7-3970X
4.3.1 Bzip2
1,17 1,17 1,175
1,165
1,18 1,185
1,175
1,011 1,021,03 1,041,05 1,061,07 1,081,091,1 1,111,12 1,131,14 1,151,16 1,171,18 1,191,2 1,211,22 1,231,24 1,251,26 1,271,28 1,291,3 1,311,32 1,331,34 1,351,36
Core 0 Core 1 Core 2 Core 3 Core 4 Core 5 Average all Cores No
HT
bzip
Core 0 Core 1 Core 2 Core 3 Core 4 Core 5 Average all Cores No HT
Figure 15: Bzip Characterization for i7-3970X
The diagram above describes the behavior of the microprocessor chip when being undervolted and one or more instances of the Bzip2 benchmark are running on one or more physical cores. The highest value of the vertical column is the starting nominal voltage which is 1.365V. The bigger the distance from the top of the diagram, the more resilient was proved to be a configuration as far as this benchmark is concerned. On the other hand, a small distance from the top implies that this configuration’s performance was poor and that it is vulnerable to the undervolting procedure. It is obvious from the diagram that the most resilient configuration is the one where a single instance of the benchmark was running on Core 3 and the rest of the processes were running on another core (more specifically on Core 5). The nominal voltage where an uncorrected error occurred was 1.165V which means that it withstood an undervolting of 195mV, which means a voltage reduction of 14.2%. The behavior of the other cores is more or less the same since their crash points differ only by up to 20mV. The average undervolting of these configurations is 190mV, which means a voltage reduction of 13.9%. On the other hand, the All Cores No HT configuration seems to be much more vulnerable. More specifically, it only withstood an undervolting of 115 mV which means a voltage reduction of only 8% in contrast to the 14.2% observed on Core 3 and the 13.9% on average. Looking into the reduction in temperatures and power consumption we can observe that for the execution on a single core a reduction of 21.5% in temperatures and 17.5% in power consumption is achieved.
4.3.2 Mcf
1,165 1,165 1,17 1,165 1,17
1,155 1,165
1,22
1,011 1,021,03 1,041,05 1,061,07 1,081,091,1 1,111,12 1,131,14 1,151,16 1,171,18 1,191,2 1,211,22 1,231,24 1,251,26 1,271,28 1,291,3 1,311,32 1,331,34 1,351,36
Core 0 Core 1 Core 2 Core 3 Core 4 Core 5 Average all Cores No HT
mcf
Core 0 Core 1 Core 2 Core 3 Core 4 Core 5 Average all Cores No HT
Figure 16: Mcf Characterization for i7-3970X
When conducting the same tests on the cores while running Mcf the resilience of all the microprocessors seems to be greater. The more resilient core is proved to be Core 5 which indicates that each core behaves differently under different workloads since core 5 was the less resilient core regarding the Bzip2 experiment. More specifically, Core 5 withstood 205mV, a 15% reduction of voltage, before the uncorrected error occurred at 210mV. The average percentage of voltage reduction for the execution on a single physical core is 14.2% which corresponds to an average undervolting of 195 mV.
Again, the less resilient configuration is the one where the microprocessor is under maximum stress, where an undervolting of 140 mV was achieved before the system crash at a 145mV reduction from the nominal voltage. This amount of undervolting corresponds to a 10.2% reduction from the starting nominal voltage which means that there is a 5% difference between the more resilient and the less resilient configuration.
On the field of temperature and power consumption an average reduction of 21.4% is observed concerning temperature and a 15.7% reduction in power consumption.
4.3.3 Milc
When observing the results of the execution of this specific benchmark, depicted at the following diagram, it can be seen that all the configurations are less resilient compared to their resilience on Bzip2 with the exception of Core 5, which is actually more resilient strengthen our hypothesis that each core behaves differently than the others.
1,18 1,175 1,18 1,19 1,19
1,175 1,18
1,21
1,011 1,021,03 1,041,05 1,061,07 1,081,091,1 1,111,12 1,131,14 1,151,16 1,171,18 1,191,2 1,211,22 1,231,24 1,251,26 1,271,28 1,291,3 1,311,32 1,331,34 1,351,36
Core 0 Core 1 Core 2 Core 3 Core 4 Core 5 Average all Cores No
HT
milc
Core 0 Core 1 Core 2 Core 3 Core 4 Core 5 Average all Cores No HT
Figure 17: Milc Characterization for i7-3970X
Core 5 is proven to be the most resilient configuration while on the other hand the All Cores No HT configuration is again the least resilient. However, their difference is not as great as on the other two benchmarks. Their difference is only 2.6% from the starting nominal voltage. More specifically, the reduction that Core 5 withstood is 13.5% while the reduction withstood by the All Cores No HT configuration is 10.9%. Α reduction of temperatures by 18.3% was observed when running the benchmark on a single core, while power consumption was reduced by 14.5%.
4.3.4 Namd
No uniformity of the results can be observed on this benchmark’s experiments. The All Cores No HT configuration, which is the least resilient, withstood an 11.3% reduction from the starting nominal voltage. When compared to the most resilient core for this benchmark, which is core 0, the difference is 3.4%, since core 0 withstood an undervolting of 200mV (14.7% of the starting voltage) before an uncorrected error occurred. When compared with the less resilient core, the All Cores No HT configuration is surpassed by only 15mV (core 4). The undervolting withstood is only 12.4% lower than the starting nominal voltage. We can easily see that the All Cores No HT configuration and Core 4 differ by only 1.1%. Concerning the power consumption, an average of 17.5% reduction was observed on the executions of the benchmark at the lowest voltage before the system crashed when compared to the executions at the starting nominal voltage. The corresponding temperature values were reduced by 22%.
1,16 1,17
1,185
1,165
1,19
1,17 1,175
1,205
1,011 1,021,03 1,041,05 1,061,07 1,081,091,1 1,111,12 1,131,14 1,151,16 1,171,18 1,191,2 1,211,22 1,231,24 1,251,26 1,271,28 1,291,3 1,311,32 1,331,34 1,351,36
Core 0 Core 1 Core 2 Core 3 Core 4 Core 5 Average all Cores No
HT
namd
Core 0 Core 1 Core 2 Core 3 Core 4 Core 5 Average all Cores No HT
Figure 18: Namd Characterization for i7-3970X
4.3.5 Hmmer
1,17
1,195
1,185 1,185 1,19 1,195
1,185
1,23
1,011 1,021,03 1,041,05 1,061,07 1,081,091,1 1,111,12 1,131,14 1,151,16 1,171,18 1,191,2 1,211,22 1,231,24 1,251,26 1,271,28 1,291,3 1,311,32 1,331,34 1,351,36
Core 0 Core 1 Core 2 Core 3 Core 4 Core 5 Average all Cores No HT
hmmer
Core 0 Core 1 Core 2 Core 3 Core 4 Core 5 Average all Cores No HT
Figure 19: Hmmer Characterization for i7-3970X
As it can be observed by the diagram above there is a significant difference between the most resilient and the least resilient configuration. As with all the previous benchmarks, the least resilient is the All Cores No HT configuration where an instance of the benchmark is executed simultaneously on all the physical cores and after each execution the voltage is reduced by 5mV until the system crashes due to the occurrence
of an uncorrected error. More specifically, this difference is 60 mV in absolute numbers and 4.4% relative to the starting nominal voltage which is 1.365V. The most resilient core withstood a 13.9% reduction in voltage while the least resilient cores (cores 0 and 5) withstood a 12% reduction. Concerning the average temperature reduction for the execution of the benchmark on a single physical core which is operating at the lowest safe voltage level, it was found that it is 18.8 % while the corresponding power consumption reduction is 16%.
4.3.6 H264ref
1,18 1,18 1,175 1,185 1,19
1,17 1,18
1,21
1 1,01 1,02 1,03 1,04 1,05 1,061,07 1,08 1,091,1 1,11 1,121,13 1,14 1,15 1,16 1,17 1,18 1,191,2 1,21 1,22 1,23 1,24 1,251,26 1,27 1,28 1,291,3 1,31 1,321,33 1,34 1,35 1,36
Core 0 Core 1 Core 2 Core 3 Core 4 Core 5 Average all Cores No HT
h264ref
Core 0 Core 1 Core 2 Core 3 Core 4 Core 5 Average all Cores No HT
Figure 20: H264ref Characterization for i7-3970X
As can be seen by the diagram above, core 5 is the most resilient while on the other hand core 4 is the least resilient. It is obvious once again that the least resilient configuration is the one where all the physical cores of the microprocessor are executing an instance of the H264ref benchmark. Core 5 withstood a reduction of 190mV (13.9%) before the uncorrected error occurred. On the other hand, core 4 withstood for 170mV (12.5%) which in relative numbers is a 1.4% reduction from the 1.365V nominal voltage. As far as the least resilient configuration is concerned it is proved that it withstood a reduction of 10.9% (150mV) of the nominal voltage. As far as the power consumption is concerned, H264ref is proved to have on average 17.7%
lower power consumption when the microprocessor is operating at a lowest safe voltage level. With the same, the reduction of temperature is on average 21.1%.
4.3.7 Gobmk
1,175
1,165 1,175
1,19 1,18 1,185
1,175
1,23
1,011 1,02 1,031,04 1,051,06 1,07 1,081,091,1 1,11 1,12 1,131,14 1,15 1,161,17 1,181,191,2 1,211,22 1,23 1,241,25 1,261,27 1,28 1,291,3 1,311,32 1,33 1,341,35 1,36
Core 0 Core 1 Core 2 Core 3 Core 4 Core 5 Average all Cores No HT
gobmk
Core 0 Core 1 Core 2 Core 3 Core 4 Core 5 Average all Cores No HT
Figure 21: Gobmk Characterization for i7-3970X
When running this benchmark, one can observe the great gains of undervolting which can be translated into gains in frequency or energy consumption. More specifically, Core 1 is proved to be the more resilient among the cores by withstanding a 195mV voltage reduction from the nominal voltage (1.365V). This means a reduction of 14.2% while on the other hand Core 3 proved to be the weakest, withstanding a reduction of 12.5%. Moreover, the configuration where each core executes one instance of the benchmark simultaneously is proved to be very weak since it only withstood a 130mV reduction (9.5%). The execution of the benchmark individually on one core operating at the lowest safe voltage showed an average reduction of 20.4% in temperature and by 16% in power consumption.
4.3.8 Zeusmp
This is the only benchmark whose resilience to undervolting when the microprocessor is fully stressed can be directly compared to the resilience of the other configurations. The difference between this configuration and the least resilient core is only 5mV while the difference with the most resilient one is 30mV (2%). The maximum reduction in voltage that was withstood was a reduction of 195mV (14.2%) that appeared on both cores 3 and 5. On the other hand, the least resilient core is core 0 that withstood a voltage reduction of 170mV (12.5%) before the system crash. Last but not least, a respective reduction of 21.3% and 17% was observed while running the benchmark on a single physical core which has reached the lowest safe voltage level regarding temperature and power consumption.
1,19
1,175 1,17 1,165 1,175
1,165 1,175
1,195
1,011 1,021,03 1,041,05 1,061,07 1,08 1,091,1 1,111,12 1,131,14 1,151,16 1,171,18 1,191,2 1,211,22 1,231,24 1,251,26 1,27 1,281,291,3 1,311,32 1,331,34 1,351,36
Core 0 Core 1 Core 2 Core 3 Core 4 Core 5 Average all Cores No HT
zeusmp
Core 0 Core 1 Core 2 Core 3 Core 4 Core 5 Average all Cores No HT
Figure 22: Zeusmp Characterization for i7-3970X
4.3.9 Conclusions for i7-3970X
To conclude, the benchmarks can be divided into two categories. In the first category belong the benchmarks that are proven to be less resilient to the undervolting such as Bzip2, H264ref, Hmmer, Gobmk and Milc. These benchmarks, when executed on a single physical core can withstand undervolting of around 180mV while on the other hand the benchmarks that belong in the other category such as Zeusmp, Namd and Mcf can operate without uncorrected errors at greater reductions of voltage (up to 190 or even 195mV) and therefore the gains in energy consumption are greater. Also in the second category, the configuration of full CPU stressed workloads is also more resilient compared to the resiliency of the same configuration for the benchmarks of the first category.
Benchmark
Configurations Core
0
Core 1
Core 2
Core 3
Core 4
Core 5
Benchmark Average by
Cores
All Cores No HT
Bzip2 195 195 190 200 185 180 191 115
Mcf 200 200 195 200 195 210 200 145
Milc 185 190 185 175 175 190 183 155
Namd 205 195 180 200 175 195 192 160
Hmmer 195 170 180 180 175 170 178 135
H264ref 185 185 190 180 175 195 185 155
Gobmk 190 200 190 175 185 180 187 135
Zeusmp 175 190 195 200 190 200 192 170
Core Average
by Benchmarks 191 191 188 189 182 188
Table 23: Crash Points (as offset from starting nominal voltage) for all Benchmarks for all configurations
Benchmark
Configuration
Core 0 Core 1 Core 2 Core 3 Core 4 Core 5
Benchmark Average by
Cores
All Cores
No HT
Bzip 22.7% 21.7% 21.6% 24.6% 18.7% 22.3% 21.9% 10.8%
Mcf 22.6% 20.2% 17.5% 22.9% 23.2% 22.3% 21.5% 9.7%
Milc 18.7% 20% 17% 20% 19.4% 19.7% 19.1% 9.5%
Gobmk 21.2% 20.7% 18% 20.8% 21.7% 20.9% 20.6% 8.2%
Zeusmp 20% 20.7% 18.5% 23.4% 23.1% 22.6% 21.4% 9.5%
Namd 23.1% 23.4% 17.2% 23.7% 22.2% 23.1% 22.1% 9.3%
Hmmer 20.7% 17.6% 18.6% 21.1% 18.8% 16.2% 18.8% 8%
h264ref 23.8% 21.9% 20.9% 20.9% 18.2% 21.1% 21.1% 11.3%
Core Average by Benchmarks 21.6% 20.8% 18.7% 22.2% 20.7% 21%
Table 24: Highest Temperature gains for all Benchmarks on all Configurations for i7-3970X
Benchmark
Configuration
Core 0 Core 1 Core 2 Core 3 Core4 Core 5
Benchm ark Average
by Cores
All Cores No HT
Bzip 19.6% 17.1% 17.4% 17.5% 15.4% 17.3% 17.4% 0%
Mcf 16.6% 13.8% 14.5% 15.9% 15.2% 15.6% 15.3% 3%
Milc 13.8% 14.7% 13.5% 14.8% 11.6% 15.6% 14% 0%
Gobmk 16% 16.6% 16.3% 16.2% 16.3% 16.8% 16.4% 0%
Zeusmp 17.1% 17.9% 16.1% 18.2% 18% 14.5% 17% 0%
Namd 18.4% 15.8% 14.5% 17.8% 17.4% 19.2% 17.1% 0%
Hmmer 16.5% 16.6% 15.8% 19.4% 16.6% 15.4% 16.7% 0%
h264ref 20.3% 16.5% 18.3% 14.1% 15.3% 19% 17.3% 0%
Core Average by
Benchmarks 17.3% 16.1% 15.8% 16.7% 15.7% 16.7%
Table 25: Highest Power Consumption gains for all Benchmarks on all Configurations for i7-3970X