Originally Posted by The Stilt
All Kaveri / Godavari parts are able to run at 2400MHz MEMCLK as long as the settings are correct.
It should require no additional VDDNB voltage either, unless your chip is running extremely hot. To ensure stability in all conditions I would recommend using the default voltage together with 80% or 60% load-line adjustment.
On boards which feature an advanced VRM controller (IR356x) it is important NOT to use the fixed voltage mode.
This is because the output voltage feeds two different domains, with two different voltage level requirements; CNB has default voltage < 1.1250V while the GNB always defaults to 1.225V (KV) or 1.275V (GV) in 3D. If you enable the fixed voltage mode and set it to eg. 1.150V you will most likely have no issues on desktop, but in 3D the GPU might potentially crash due undervoltage. Likewise if you set the fixed voltage to e.g. 1.2750V, you will just end up wasting power on the desktop when the GPU is not used. When the fixed voltage option is used the VID commands issued by any of the domains (CPU, CNB, GNB) are completely ignored. As the name indicates, the voltages are fully fixed.
When using unofficially supported MEMCLK (i.e. 2400MHz) most ODMs automatically add a voltage offset to the VDDNB voltage. In most cases the automatic voltage offset is way too large and it only results in excessively high power consumption with no additional stability. In fact in some cases the automatic voltage offset is so high that the resulting output voltage violates AMD´s electrical specifications, and potentially can cause instability or even degradation. To ensure there is no excessively high automatic voltage offset added by the bios, you must ensure that the offset voltage is NOT set to Auto. If you are adjusting either CPU, NCLK, GPU or MEMCLK frequencies above their default level you should use at least 0.00625V offset voltage (±, doesn´t matter). Setting the offset value to anything other than Auto will prevent the bios from using the automatic voltage offsets.
Overclocking the APUs has always been extremely tricky since adjustments made to one domain also affect the other.
Kaveri / Godavari will not be able reach MEMCLKs higher than 2400MHz, and the reason for this limitation is rather simple.
During the boot the memory controllers perform a calibration process dubbed as training. The training process will adjust the settings until the signal patterns match the predefined pattern for the set memory frequency. The memory controllers perform training at several different frequencies depending on the desired memory clocks. For example when the desired MEMCLK is 2133MHz the memory training will be performed at 800, 1066, 1333, 1600, 1866 and 2133MHz frequencies. The memory signals can drift from their trained settings only so slightly without causing any problems. If the actual signaling changes to either direction, at some point it will either result in memory errors or system crash depending on how much apart the signaling is and how many errors are produced because of it.
This is the reason why using 110MHz BCLK with 3:28 (1866MHz) MEMCLK ratio (= 2053MHz) might be unstable while the 3:32 (2133MHz) MEMCLK ratio at 100MHz BCLK works just fine.
The memory controller in Kaveri / Godavari cannot train the memory for any frequency higher than 2400MHz.
You can try to reach higher frequencies by increasing the BCLK frequency from 100MHz, however in that case you are playing with the signal margins which is never advised for the sake of stability.
Regardless how much you increase the BCLK above the default 100MHz value, the memory will not be trained for any other frequency but 2400MHz. That´s why anything higher than 2400MHz will not be reliable or stable, no matter how much you fool around with the voltages or timings.
These limitations will apply regardless of the motherboard as the limitation is purely in the memory controller.
Some motherboards might have slightly shorter trace lengths, larger trace spacing or cleaner supply for VDDIO or the reference voltages, but the difference between the designs will still be non-existent due the limitations set by the memory controller itself.
To maximize the memory performance and stability on Kaveri / Godavari, use following configuration:
- Two DRAM modules
- Dual rank modules (two sided with 8 ICs on each side)
- Hynix CFR (2Gb), MFR (4Gb) or Samsung D-Die (2Gb) or Q-Die (4Gb)
- Micron / Elpida and Nanya ICs are usually good for 1600-1866MHz on AMD
The penalty from using single rank modules is worth around one speed class or 133MHz (2400MHz SR = 2133MHz DR, 2133MHz SR = 1866MHz DR).
On average (depending on leakage) the GPU reaches it´s maximal frequency with 1.27 - 1.30V actual.
In most cases the GPU and NCLK frequency scaling starts to show retardation if the voltage is raised beyond that.
The CNB/GNB voltage scaling is rather poor in Kaveri / Godavari, so trying to reach frequencies beyond the linear scaling curve by using excessively high voltage is futile and after a certain point dangerous to the silicon. Unless the iGPU is (for some odd reason) used for computing, overclocking the iGPU is rather pointless as at 2400MHz MEMCLK there is not enough of bandwidth to keep the GPU fully fed even at 600MHz.
The breakdown (absolute maximum) voltage for Kaveri / Godavari CNB/GNB (VDD_NB) is 1.320V.