Why has the CPU clock frequency not increased in the last 5 years?

Original link: Quora translation: Jay Weibo-Sophy
Link: Http://www.jointforce.com/jfperiodical/news/show/321?m=d03 can also earn money on the Jointforce platform package ~

The CPU clock frequency has not increased in the last 5 years is caused by many different categories of reasons.
First, the Power

The graph is the relationship between the clock frequency and the power consumption

When designing a CPU's microarchitecture, one of the key design decisions is how to achieve higher performance. In the Pentium 4 era, Intel chose to have a very high clock frequency and a relatively narrow pipeline. This approach has many advantages, one of which is that it is easy to speed up single-threaded and serial code. There is no need to operate many parallel instructions within the software, so most software will see its benefits immediately.

However, this approach also has its drawbacks, ignoring the execution flaws of the Pentium 4 itself. Here, we only talk about the flaws in the concept itself. The main thing is that the micro-architecture of the CPU has been in conflict with the wall, and the high-frequency microarchitecture is not suitable for many low-power design techniques that have been invented to handle power problems. Here, I'll touch on two major low-power design approaches.

One is the clock gating technology, the clock gating technology will insert a clock start system before each State element (register, lock, etc.), so that if no new data is written, the component will have no clock control. This saves a lot of wasted charge/discharge time due to the same cache being written back. This method also inserts an additional delay (gated function) into the clock path. High-frequency design is generally low-margin operation and is not suitable for inserting additional variable delay in the most critical signal (clock).

Another common technique is power gating. This involves placing transistors on voltage sources on different parts of the chip. Typically, when not in use, those different function modules and power supply sections are turned off, but high-frequency designs do not often do so. The power control transistor not only needs to add an additional pressure drop, which slows down the transistor switching speed, and a very thin pipelined processor simply does not have many parts that can be shut down at any given time.

Therefore, from the micro-architecture point of view, high-frequency and thin design is not just intelligent power-wise.

Second, transistor scaling

The other main reason why processor frequency does not rise is simply that the transistor itself is not getting faster.

Others refer to the size of the transistor's width, but the width of the transistor is actually steadily declining and will continue, and Moore's Law works well in this respect.

Intel is currently manufacturing 32 nm of HKMG (high-k insulating layer + metal gate) on a 45 nm basis. Two years ago, it was 65 nanometers, before it was 90 nanometers. TSMC, IBM and GlobalFoundries began production of 28 nano-chips this year. Intel is planning to adjust to 22 nm. (Update: 14NM has come out).

The problem, however, is that they are not getting faster and quicker as the size of the transistors grows. To understand this, a bit of MOSFET (golden oxygen Half-effect transistor) background is necessary.

It is well known that the switching speed of transistors depends on many factors. One of the main factors is the strength of the electric field created in the gate (control to switch). The intensity of the electric field depends on the two areas of the gate (which is smaller, the transistor shrinks), and the gate thickness.

As the transistor shrinks, the area of the door decreases. In the past, the reduction in the area of the brake poles meant that the gates of a transistor could be made thinner. If you know how a basic capacitor works, you know that the smaller the distance between the two conductive plates, the stronger the electric field between them. This working principle is equally common on MOSFETs. The thinner gate dielectric causes a stronger electric field to pass through the transistor channel, which means that the transistor switches faster. Reduced transistor gate size means that the gate can be made thinner, and the load capacitance increases harmless.

However, as for 45 nm, now the brake medium is about 0.9 nanometers thick-about a silica molecule size, so it is impossible to make a thinner. As a result, Intel replaced silica with High-k, a material based on hafnium materials, and became a brake-electrode dielectric (many people suspected of being hafnium silicate). They also turned the material from polysilicon to metal.

This method helps to increase the speed of the transistor, but it is too expensive and can only be a stopgap. The simple thing is that every time we use the simple scaling that we've already had to reduce the transistor, it will lead to a faster transistor end.

Three, Chip scaling

Another major reason for the slowdown in frequency ramp is that transistors are no longer unique-in some cases, even the largest-the key to how fast a processor can run. The wires connecting these transistors are now a major factor in the delay.

As transistors become smaller, the wires that connect them become thinner. Thin lines mean higher resistance and lower currents. The fact is that smaller transistors can drive a small amount of current, and it is easy to see that the switching speed of the transistor only partially determines the path delay of the circuit.

Of course, many techniques can be used to deal with this problem in the chip design process. An engineer with good layout and cabling will try to plan the route of its clock and data signal in a similar way, so that two signals can be transmitted simultaneously and at the same time to reach their destination. For data-intensive chips, lightweight design can be a very effective solution, such as a fixed-function video codec engine or network processor.

However, a microprocessor with web interaction is a very complex, unconventional design, where data access to multiple locations does not always follow the clock rules, it has feedback paths and loops, and has centralized resources such as risk tracking, scheduling, branch prediction, register files, and so on. In addition, the heavy-control design can easily be copied to more cores, but the required thin lines are complex to improve the processor frequency by standard methods.

[Reprint please keep the source, translator and translation links.] ]

Why has the CPU clock frequency not increased in the last 5 years?