Anatomy of a Power Delivery Network
As Figure 1 illustrates, power delivery is a full system problem that starts all the way at the main power supply and extends to a power grid in the processor that ultimately reaches the transistors that perform computations on die. For a desktop, the power supply will convert from 110V or 220V AC to a 12V DC current that is distributed across the motherboard to the processor and other components. In a notebook or smartphone, things will be a bit different; typical lithium-ion batteries output a 3.7V DC current so there is no AC to DC power conversion and the conversion ratio is much lower.
For standard processors, such as those from AMD, the voltage regulation modules (VRMs) convert down to about 1V. Generally, the VRMs are placed in close proximity to the processor so that most of the distance that the power travels will be using the 12V signals on the motherboard. The 1V power supply is transmitted a short distance across the motherboard, through the processor package and into the processor itself through a set of bumps. The processor contains a power grid that fans out from the bumps and uses the various metal interconnect layers to deliver power to the transistors on die. Motherboard voltage regulators are fairly slow and operate at around 1MHz, meaning that the VRM can only adjust the output voltage every microsecond.
Many Intel-based systems follow the same principles, but with an extra stage in the power delivery. The FIVR or fully-integrated voltage regulator is integrated into the processor die itself and fans out to dozens of power rails on different blocks (e.g., CPU cores, CPU L2 caches, different GPU blocks, etc.). The FIVR is used in most server processors starting with the Haswell generation. It is also used in the Haswell and Broadwell client processors and now Ice Lake and Tiger Lake client. Note that the Skylake client family (including Coffee Lake, Comet Lake, etc.) do not use the FIVR. In these systems, the motherboard VRMs convert the 12V (or 48V) signal down to about 1.8V, which is transmitted from the VRMs, across the motherboard, the package, and the processor pins and into the FIVR. The FIVR is responsible for the last stage of power conversion and finally converts the voltage from around 1.8V down to around 1V, depending on the needs of the particular power rail.
One advantage of the FIVR is that the voltage delivered from the motherboard VRMs to the processor is about twice as high as in a conventional system. Using a higher voltage reduces the necessary current by a similar factor of two, reduces the number of power pins, and boosts efficiency. The downside is that voltage conversion is never 100% efficient and the FIVR does lose some efficiency as a result. The relationship between the transmission efficiency gains and conversion efficiency losses is highly dependent on the exact situation. Overall for high-power processors, the FIVR appears to be a win. Additionally, the FIVR is amazingly fast – it operates at 140MHz, two orders of magnitude faster than a motherboard VRM.