Performance vs Power: Overclocking the Raspberry Pi

Preparation

Raspberry Pi 2 Model B.
A power meter bought on Taobao for 37 RMB.
Power supply from my Nubia Z9 Max, 5V 1.5A.

Testing Method

Modify config.txt to overclock via arm_freq.
Change scaling_governor from powersave to performance to ensure the CPU runs at maximum frequency.
Use aircrack-ng -S to measure power consumption and performance under full load.

Data Table

Frequency (MHz)	Idle Power (W)	Full Load Power (W)	Performance (k/s)	Performance/Freq	Performance/Power
600	1.75	2.31	339	0.56500	146.75
700	1.82	2.65	397	0.56714	149.81
800	1.83	2.81	455	0.56875	161.92
900	1.84	2.94	514	0.57111	174.83
1000	1.83	3.08	571	0.57100	185.39
1050	1.84	3.13	600	0.57143	191.69
1100	1.89	3.37	629	0.57182	186.65

Note: At 1100MHz, the system could not run stably under full load at default voltage. After testing, setting over_voltage=2 made it stable.

Conclusion

With governor set to powersave, idle power consumption is 1.75W.
With governor set to performance, idle power consumption remains almost unchanged across frequencies, around 1.84W.
Performance per frequency slightly improves as frequency increases.
With over_voltage=0, performance per watt improves as frequency increases.

Therefore, to improve performance per watt, we can overclock as much as possible without increasing voltage, provided the system remains stable.

Moreover, even with overvolting at 1100MHz, its performance per watt is still better than 1000MHz at default voltage. Methods to improve low-frequency performance per watt are discussed in the postscript.

Postscript

These are experiments I did after completing the experiments above.

Unexpected Findings

After overclocking, even when the CPU frequency stays at 600MHz, there is a slight increase in power consumption and performance under full load.
At maximum frequency 1000MHz but actual frequency 600MHz, full load power is 2.33W, performance is 341k/s. Idle power remains unchanged (original idle power 2.31W, performance 339k/s).

This indicates that overclocking slightly increases full-load power consumption and performance even at the lowest frequency (600MHz).

Stability Testing with Overvolting

1
2
arm_freq=1050
over_voltage=3

No difference in idle power consumption with powersave governor.
With performance governor, idle power is 1.91W, full load power is 3.38W, performance is 599k/s.
At the same frequency, overvolting increases power consumption, does not improve performance, causes more heat, and reduces performance per watt.

Thus, if the system is stable, overvolting is not recommended.

Methods to Improve Performance per Watt

An unexpected discovery: I set over_voltage to a negative value, hoping to undervolt. I thought it wouldn’t boot, but surprisingly it worked! This means we can improve performance per watt by lowering the voltage.

1
2
arm_freq=800
over_voltage=-3

With performance governor, idle power is 1.78W, full load power is 2.63W, performance is 456k/s, performance per watt = 173.38 (original idle power 1.83W, full load power 2.81W, performance 455k/s).

This suggests that as long as the system remains stable, we can lower the voltage to improve performance per watt. (In fact, my laptop i7-4710MQ CPU has been undervolted by 0.1V for a long time.)

Therefore, if you find your Raspberry Pi’s performance sufficient (i.e., you’ve reached your target frequency), you can reduce the voltage as much as possible while keeping the system stable to improve performance per watt.