I can set CPU at 2Ghz now, but it will not increase performance.
Because L_LV1_volt = 1080 is the same value for 1.8Ghz which formerly used.
When I do Geekbench test, it jump back to 1.8Ghz or lower
So, if we want ârealâ 2Ghz, we need to give L_LV1_volt a little higher value than 1080.
But what will be the safe value to do the real overclock but not burn BPI-M3?
Thatâs not the reason. There are certain aspects and some of them are related. You need higher VDD_CPUX at higher clockspeeds. The higher VDD_CPUX --> the higher also the heat emissions. Since A83T is prone to overheating: the higher VDD_CPUX --> earlier throttling.
So without understanding the relationship between clockspeed, VDD_CPUX (high enough to ensure reliability, low enough to not increase heat too much) and temperatures you canât tune anything. And without something like RPi-Monitor drawing nice graphs youâre lost anyway:
TL;DR: BPi M3 can not be âoverclockedâ without liquid cooling. Without a heatsink this device is limited to 1.2GHz, with heatsink and much airflow 1.6Ghz is possible. Everything known since 1st devices have been delivered half a year ago:
The only âoverclocking modeâ that works is the absolute moronic one: 2.0 GHz when the device is idle, but way below when itâs busy and higher clockspeeds would be needed (thatâs what you get with all the settings above)
Absolute maximum ratings for VDD_CPUX are 1300mV (for example: For H3 the datasheet claims that the absolute maximum is 1.5V but VDD_CPUX should not exceed 1400mV under normal conditions).
Based on that one could assume that getting close to 1200mV should be ok on A83T. But thatâs just absurd since the SoC already starts to insanely overheat already when fed with 1000mV. Then itâs quite normal that you get tested/recommended values not from the datasheet but from comments in Allwinnerâs BSP. In our case that means:
In other words: Allwinner says feeding just 1000mV should be ok up to 1800MHz and 1080mV should be used above
And now comes the interesting part: How to verify at which voltage / clockspeed combinations A83T works reliable? Thatâs easy up to 1600MHz since here you can test the whole thing under âworst case conditionsâ that are necessary since otherwise youâre testing wrong. But exceeding these 1600MHz will be hard since when you do serious testing the SoC will either throttle down or you have to increase the trip points accordingly: Please keep in mind that throttling starts when the internal thermal sensor reaches just 65°C:
So to test higher clockspeeds, the trip points have to be increased a lot. Then to prevent throttling jumping in too early the dvfs operating points have to be decreased step by step until reliability is concerd (bit flips being reported). We (linux-sunxi community) went through this process for A83Tâs siblings H3 and A64 already. And anyone interested in better performance with BPi M3 can do the same. But when you do testing seriously then you will waste a lot of time and will end up with the conclusion that A83T is a â1.6GHz SoCâ since higher clockspeeds would require liquid cooling or annoying fans.
And the most important thing to understand: Overclocking is a nice try when you want to do numbercrunching on an SBC (which is somewhat weird though) but itâs soooo useless when itâs about real world performance since there other settings and performance metrics matter way more. So by being focused on higher CPU clockspeeds instead of things like random IO performance, IRQ distribution accross different CPU cores, improved scheduler and kernel settings you simply gain nothing.
I notice throttling during Geekbench testing. That seems normal in all SBC
Comparing with my Odroid XU4 which is based Samsung Exynos5422. A83T CPU speed is really not a big concern. But the LDDR3 speed is very slow in BPI-M3. it drags the whole benchmark down!
