Based my own tests on several different types of "Godavari silicon" the 7870K´s do indeed overclock better than Kaveri.
For the CPU part the average difference is around 300MHz, 200MHz for the CNB part and around 100MHz for the GPU.
The absolute clock ceiling for the CPU will be around the same (< 4800MHz) as it is process limited but Godavari will be able to do the same clocks as Kaveri with significantly less voltage.
The average clocks you can expect with a high-end motherboard and cooling are up to 4500 - 4700MHz for the CPU, 2100 - 2200MHz for CNB and 1040 - 1114MHz for the GPU. Compatible (Samsung or Hynix based) memory should run stable at 2400MHz without even touching the UNB voltage (CNB/GNB).
Both Kaveri and Godavari are extremely sensitive to fluctuations in CPU core voltage.
Since Godavari clocks higher than Kaveri it is even more sensitive to the fluctuations as the tolerance decreases at higher frequencies.
For that reason the load-line must be calibrated extremely accurately. At higher clocks you want the core voltage to stay within 15mV of the set-point even during the maximum loadstep. Both undershoot (droop) and overshoot hurt equally so simply maxing out the load-line control is not advised. Usually the -60% setting works the best.
Due the sensitivity to both voltage stability and other voltage characteristics the VRM plays a huge role when these parts are overclocked.
On a high-end motherboard with a digital VRM on it a chip might do 4.7GHz with just 1.45V. On a different lower-end motherboard the same chip might require 1.5V for 4.5GHz and be unable to reach any higher in otherwise identical conditions.
The CPU frequency usually reaches it´s peak at 1.45V - 1.475V depending on leakage.
It will accept higher voltage but usually the returns are non-existent. There is absolutely no point in trying to force the chip to do higher than xMHz if it requires more than 1.475V (actual).
The default voltages on 7870K´s are extremely high but also extremely conservative at the same time.
This means that there is a huge amount of undervolting potential in these chips. Not just for the CPU but for the UNB also.
On average the CPU requires around 200mV less voltage (abs. 1.2 - 1.25V) for the base frequency of 3.9GHz than the factory values indicate.
For the UNB the number is around 125mV (abs. 1.150V)
The power savings from undervolting are quite impressive.
The measurements are DCR (current over inductors) meaning they represent the actual power consumed by the APU.
Stock vs. Tuned voltages:
Pb0 (4100MHz) = 1.43750V vs. 1.1960V
P0 (3900MHz) = 1.38750V vs. 1.1530V
P1 (3800MHz) = 1.35000V vs. 1.1180V
P2 (3700MHz) = 1.30000V vs. 1.0960V
P3 (3500MHz) = 1.23750V vs. 1.0640V
P4 (3000MHz) = 1.08750V vs. 0.9630V
GNB (CNB + GNB) during 3D = 1.27500V vs. 1.0480VTest: 3DMark11 Performance
Stock PMax VDD_CR DCR = 52.125W
Stock PMax VDDNB_CR DCR = 47.6875W
Total = 99.8125W
UV PMax VDD_CR DCR = 36.250W (-30.46%)
UV PMax VDDNB_CR DCR = 35.625W (-25.3%)
Total = 71.875W (-28%)Test: 3DMark Sky Diver
Stock PMax VDD_CR DCR = 55.125W
Stock PMax VDDNB_CR DCR = 48.625W
Total = 103.75W
UV PMax VDD_CR DCR = 36.8125W (-33.21%)
UV PMax VDDNB_CR DCR = 37.1875W (-23.52%)
Total = 74.0W (-28.7%)Test: 3DMark Fire Strike
Stock PMax VDD_CR DCR = 47.750W
Stock PMax VDDNB_CR DCR = 46.760W
Total = 94.51W
UV PMax VDD_CR DCR = 32.3125W (-32.33%)
UV PMax VDDNB_CR DCR = 35.000W (-25.15%)
Total = 67.3125W (-28.8%)Test: Cinebench R15
Stock PMax VDD_CR DCR = 61.0625W
UV Pmax VDD_CR DCR = 35.6875W (-41.6%)