Power Delivery in a Modern Processor

Pages: 1 2 3 4 5

Decoupling Capacitors Smoothly Deliver Power

To address the mismatch between nearly instantaneous transient current spikes and the latency of voltage regulators, modern systems rely on decoupling or bypass capacitors. These capacitors store energy and can quickly provide power to ensure a consistent voltage while the regulators are just beginning to respond. Returning to Figure 1, systems include decoupling capacitors (or decaps) at every step of the power delivery network. Motherboards typically include capacitors in many places, but especially around the socket as illustrated in Figure 2. Processor packages also incorporate decoupling capacitors, typically around the edges and on the underside. Lastly, processors use a variety of on-die capacitors; these are the closest to the active circuits and provide the fastest response times to transients.

Decoupling capacitors

Figure 2. Decoupling capacitors around processor socket.

On-die capacitors are available in a variety of flavors. The simplest type of on-die capacitor is a regular transistor, sometimes called a MOSFET capacitor. These decaps can be easily inserted in standard cells very close to the crucial areas that are expected to have high switching noise. Since the decaps are close to the switching activity, they can easily absorb the noise and quickly supply additional current.

Additionally, chips that are designed using automated tools tend to have whitespace, or areas that are empty due to the imperfect nature of the tools and constraints on placing differently shaped blocks close together. Filling up whitespace with decaps is a fairly common technique, since it is conceptually free. While MOSFET capacitors are available in any digital process technology and are simple to place, they are not ideal capacitors. Like any other transistors, they leak and also can be challenging to fit in areas of a chip that are very congested. Another approach is to modify the process technology and create more specialized structures such as metal-insulator-metal (MIM) capacitors, metal-oxide-metal (MOM) capacitors, or deep trench capacitors.

Intel 22nm MIM capacitors

Figure 3. Intel 22nm MIM capacitors for eDRAM.

As the name implies, a MIM capacitor is formed by two parallel metal layers with an insulating high-k dielectric between them. Intel’s 22nm process family includes two different types of MIM capacitors. As illustrated in Figure 3, the first type of MIM capacitor is used for a bit-cell in eDRAM and is formed in the lower M2-M4 metal layers. The second is featured in Intel’s 22FFL process and uses the extremely thick 4μm top layers for the parallel metal layers. Intel is hardly unique here – other manufacturers also offer MIM capacitors. For example, AMD employed an upper level MIM capacitor in the Zen CCX for decoupling to reduce voltage droop. MIM capacitors are generally superior to MOSFET decaps, but they are slightly farther away because they are in upper metal layers and the extra formation steps increase the cost of manufacturing a bit. MOM capacitors use the same general idea of parallel metal lines, but turned 90°. The metal lines are formed horizontally within two adjacent vertical metal layers (e.g., M3 and M4), with the interlayer dielectric oxide acting as the insulator.

Deep trench capacitors are another option, but are fairly uncommon in a logic process, because etching out the high aspect ratio trenches adds significant cost to the process. Deep trench capacitors were used for several generations of processor technology starting with IBM’s 32nm SOI logic process and continuing down to the 14nm SOI process. IBM’s deep trench capacitors are used for large eDRAM arrays, which implement various L2, L3, and L4 caches of for POWER and zArch processors, and they are also used for decoupling as well. As an example, IBM claims they were able to eliminate all package capacitors in the 32nm IBM mainframe z12 processor, replacing them with on-die deep trench capacitors. More recently at IEDM 2019, TSMC described forming deep trench capacitors in a silicon interposer. This is a clever and elegant approach, although these capacitors are not as close to the active logic as an on-die solution and therefore cannot completely replace on-die decoupling capacitors.

Pages: « Prev   1 2 3 4 5   Next »

Discuss (38 comments)