3 – VRM: Radeon HD 5970
The Radeon HD 5970 is a dual-GPU graphics card and two VRMs are used to feed both GPUs.
Here is the diagram of a HD 5970 VRM:
The HD 5970 voltage controller is a Volterra VT1165MF. This component allows to adjust the GPU voltage up to 2.0V (and adjusting the voltage of each GPU separatly is possible in theory…). The default voltage is 1V. The HD 5970 is softmod compatible: the voltage can be changed by software.
The VT1165MF drives three power MOSFETs with a PWM signal. Each power MOSFET is materialized by a Volterra VT1157SF. On a HD 5970, each GPU is fed by a 3-phase VRM.
Each GPU of the Radeon HD 5970 requires less energy than the GPU of a Radeon HD 5870. That’s why there are only 3 phases to feed each GPU (vGPU for GPU voltage) on a Radeon HD 5970 while there are 4 phases on a Radeon HD 5870:
Radeon HD 5870 and the 4-phase VRM (source)
AMD has reserved some place to add another phase on the HD 5970. Ultra high-end Cypress-based cards like the Asus Ares (true dual-HD5870) need 4-phase VRMs per GPU.
These power components dissipate a lot of power that’s why, on modern high end graphics card like the Radeon HD 5970 (of a GTX 480), they are equipped with heatsinks like this Radeon HD 5850:
Radeon HD 5850 with Thermalright VRM-R5 heatsink (source)
The VRMs of the Radeon HD 5970 are inductor-based VRMs. This kind of VRM requires off-chip inductors.
These off-chip inductors are materialized by the Coiltronics CPLA-3-50, a 3-phase power inductor (which has been designed exclusively for use with Volterra devices). The CPLA-3-50 is a 3-phase low noise component (yes, the inductor can sometimes generates a buzzing noise depending on the amount of current that is passing through…), each phase offering an inductance of 50nH (nano-Henry).
If you discover the term of Henry for first time, just know that the Henry is the unity that allows to describe an inductor. A capacitor is described by the Farad and a resistor by Ohm. The resistor, capacitor and inductor are the three fundamental passive components of electronics. Any electronic system can be described by an arrangment of these three components.
We have forget to mention two other MOSFETs on the Radeon HD 5970 PCB:
These two MOSFETs coupled with a 2-phase shoke (CLP-2-50), and controlled by a VT1165MF, feed the GPU uncore I/O. Both phases are shared between both GPUs.
The TDP of a Radeon HD 5870 board is 188W. With a voltage of around 1V, the GPU is not the unique device that consumes power otherwise that would mean a current of around 188A for the GPU, which is not possible (maybe in the future 😉 ). Let’s suppose the GPU requires 78A for its needs (ouch! FurMark is certainly running…). That leads to 78W for the GPU and 188-78 = 110W for the rest of the board. These 110W are essentially dissipated by the VRMs (especially by the MOSFETs), memory and fans. Keep in mind that these number are only assumptions but in extreme conditions, we are not far from reality.
The MOSFTETs (VT1157SF) on Radeon HD 5000 series are now hardware protected by a dedicated chip (which is the big difference with the previous HD 4000 generation – see HERE for more details). Then if the power dissipated by the MOSFETs exceeds a limit, the protection chip throttle back the graphics card in order to return in the limits. Recent FurMark stress tests with an overcloked HD 5970 (Legion Hardware or AnandTech) have shown that the protection was fully operational. Maybe a little bit too much but that is another story.