Why is the cpu no longer increasing the main frequency but taking multi-core now?

If you remember the shocking kneeling of Intel President Barrett in front of 6,500 people in 2004, you may understand this problem better. At that time, Mr Barrett's kneeling was a sincere repentance for the mistake of "dominant frequency theory".

At that time, prescott (Pentium 4) with NetBurst architecture adopted the most advanced 90nm technology, but the power consumption of CPU with 3GHz frequency exceeded 100 watt. If the frequency exceeds 4GHz, the power consumption will be great.

Therefore, you can answer questions here precisely because power consumption (heat dissipation) restricts the improvement of the main frequency.

The end of Danard's scaling law

I'm sure you've heard of Moore's Law, which tells us that the size of transistors in a chip is decreasing, so the number of transistors in a chip can increase. Although Moore's law has been revised in recent years, it does not seem to have stopped completely.

In fact, besides Moore's Law, there is also a very important law called Dennard Scale. Generally speaking, as the size of a transistor decreases, its power consumption also decreases in proportion to its area.

Moore's law and Dennard's scaling law, two good friends, together tell us that under the condition of basically unchanged power consumption, we can continuously reduce the size of transistors and accommodate more transistors in CPU.

But Pentium 4 basically declared the end of Dennard's scaling law, because Pentium 4' s performance is only 6 times that of 486, but its power consumption is 23 times that of the latter (6 1.75)!

Well, look at the picture. With the increase of transistor area density (blue line) 16 times, the power consumption only decreases by about 4 times (purple line), and the decrease of power consumption is no longer directly proportional to the increase of chip area density.

In other words, it is no longer feasible to continue to improve performance by increasing the frequency!

Multi-core can also brush performance.

How is the performance of CPU defined? Intel put it this way:

Where f is frequency, increasing f can improve CPU performance, but this road is no longer feasible.

However, we can also upgrade IPC. IPC(instruction per clock) is the number of instructions executed per clock cycle, so we have multiple cores, two cores, twice as many as IPC, and four cores, four times as many as IPC, and the performance of CPU is improved. This is why our consumer CPU has gone from 2 cores to 4 cores, then to 8 cores, and now it has risen to 16 cores.

Anyway, Moore's law can still linger, but the Danard scale is a thing of the past. Although the technology is more and more advanced, more transistors can be installed in CPU, but due to the power consumption wall, there is no way to increase the frequency of single core. The solution is to keep more cores on the chip to improve CPU performance. Of course, not all programs can support multi-core, so this potential performance gain can not always be presented, but it is definitely getting better and better.

No matter how high the engine speed is, the speed increase is not as direct as that of the cylinder! The V 12 engine will not make 9000 revolutions, and 8000 will enter the red line.

No matter how big the speaker is, no matter how high the volume is, watching movies can't be better than 7.2 channels.

Theoretically, the higher the clock speed, that is, the higher the main frequency, the faster the CPU will run. Frequency refers to the number of periodic changes per unit time. Some instructions can be completed in one clock cycle, while others require multiple clock cycles. If the clock speed is increased to 3.2GHz, the CPU will execute 3.2 billion cycles per second.

It seems difficult for everyone to understand that increasing the main frequency will improve the performance of CPU. For example, it takes 2 seconds for you to raise your hand, so it takes 65,438+0 seconds for you to complete a gesture of raising your hand, 65,438+00 gestures for 65,438+0 seconds, and 65,438+000 gestures for 65,438+0 seconds, which improves performance. In the shortest time, tens of billions of transistors in CPU can be quickly switched on and off, and the computing power of CPU can be improved.

Increasing the main frequency of CPU can really improve the performance of CPU, but it will soon be destroyed.

In the early days, most people thought that the higher the main frequency, the higher the CPU performance, and CPU manufacturers used to guide the general public. This triggered a protracted battle between Intel and AMD.

AMD's Athlon series took the lead in breaking through 1GHz, which made Intel lose its foothold and rushed to launch Pentium 3 series. Pentium 3, which was hastily launched, still had many problems, which didn't help Intel regain a game, so Pentium 4 based on NetBurst architecture was quickly launched. Athlon appeared around 1. 1GHz, and Pentium 4 quickly pulled around 1.4GHz, which caused AMD to lose its price advantage.

Although Pentium 4 won the market, it soon found the problem. Pentium 4 is not as good as Pentium 3 in many aspects, and Pentium 3 is generally described as "high frequency and low energy".

All this is attributed to the long pipeline of NetBurst architecture to improve the main frequency. The 20-stage assembly line is not pleasant to listen to, but it is just grinding, but the pain point is that the CPU has a large calorific value, so the requirements of the CPU for fans and radiators are getting higher and higher in the later stage, which leads to the use of CPU for jiaozi and barbecue in the later stage.

AMD also made the same mistake because of insufficient performance and overclocking. It improves the frequency of CPU through an ultra-long pipeline. For example, FX-9590 has a main frequency of 4.7Ghz and a TDP of 220W, so it can't restrain air cooling, so it can only use high-end water cooling to dissipate heat. Only in this way can I have an i3 silent second stalk on the Internet, and AMD, which pursues single-core main frequency, will finally sit in the position of the second child in the Millennium. Fortunately, AMD later began to realize the seriousness of the problem, and the multi-core RYZEN series began to show signs of turning over.

It is doomed to be a joke that a single core can improve CPU performance by increasing the main frequency.

In 2004, the 64-year-old Intel CEO knelt down in front of more than 6,500 technicians and announced that he would give up the Pentium 4 with 4GHz frequency, which shows that Intel could not solve the problem of heat dissipation caused by the increase of CPU frequency. This is a turning point for Intel, and it is also a turning point from single core to multi-core, because after Sun, IBM and AMD, Intel announced that it will move towards multi-core.

The performance of CPU = clock frequency *IPC(IPC is the number of instructions completed in a clock cycle), and the power consumption of CPU is directly proportional to current * voltage * voltage * main frequency. Increasing the main frequency is likely to increase the power consumption of CPU by three times, while increasing IPC will only increase the power consumption of CPU linearly. If the IPC is increased by 1 times and the clock frequency is doubled, it is likely to produce a result, the CPU performance remains unchanged, but the power consumption is greatly reduced. Undoubtedly, multi-core can increase IPC, reduce clock frequency and improve CPU performance.

abstract

In the past 30 years, the performance of CPU has improved with the increase of main frequency, which is the cornerstone of the development of chip industry from technology, application and industry, but now the cornerstone of this building has been completely changed. It can only be said that it is too idealistic to improve the performance of CPU by increasing the main frequency of single core, so that many external factors are ignored, and the ruthless slap in the face finally makes the chip giants embark on the road of multi-core.

At present, it is not technology that restricts CPU, but heat dissipation. Intel's CPU can be relaxed to 6-7Ghz, provided that you use liquid nitrogen to dissipate heat. Considering the reality of most air-cooled heat dissipation at present, limiting the main frequency to 2-4 is also a compromise to the market. If miniature liquid nitrogen radiators are popularized one day in the future, many cores will be less important.

Welcome to reading.

First of all, it should be said that it is not that the main frequency of mobile phones is not high now, but the speed of improvement is slower than before.

Welcome to pay attention to the author and talk about technology and digital together.

Don't just use frequency to measure the single-core performance of CPU. For example, the 555Mhz Tei CPU of Milestone 1 generation can run the CPU of 1Ghz from top to bottom on HTC G7. The single-core performance of CPU can be calculated by the wheels of the vehicle. The frequency is just the speed, that is, turning multiple gears. Another factor that affects it is single-core energy efficiency, which corresponds to the diameter of the wheel. The wheel has a large diameter and does not need to rotate quickly to maintain the height. However, if the diameter is small, the speed must be increased to achieve the same speed, resulting in increased power consumption and heat generation.

Don't look at the core frequency to quantify CPU performance, look at the frequency of the core architecture. Generally speaking, the higher the frequency of the same architecture, the better the performance. For example, the performance of eight nuclear bulldozer at 3.2gHz is not as good as that of Core i5 with four cores and eight threads. The performance of Pentium N4200 with four cores and four threads is not as good as that of M5 with two cores and four threads. At present, the best performance in CPU field is Core. For example, the Jaguar architecture of the host CPU can only be similar to that of the pile driver, but far from that of Core I. Some people speculate that the performance of eight nuclear Jaguar is only a little better than that of dual-core Core i5.

With a weight of 50kg, one person can't carry it. There are two solutions. One is to exercise to increase muscle strength, and after half a year of hard training, you can basically move; Instead, I called another one to help me, and it took 1 minute to solve it. [Laughter]

What parameters can be used to measure CPU performance? I believe that many people think of CPU frequency first. Under the same architecture and technology, the higher the CPU frequency, the stronger the performance. I remember before 2003, the frequency of CPU was not small. 198 1 year, IBM's CPU frequency is 4.77Mhz, 1995, and Intel's CPU frequency reaches 100Mhz, which is more than 20 times higher.

In 2000, AMD's CPU frequency exceeded Intel 1 GHz, which increased by 10 times in these five years. Then in 2003, Intel's CPU frequency reached 3.7Ghz, and in just three years, the frequency increased several times. By 202 1, the maximum CPU frequency is 5.3GHz, which can be gradually improved compared with the past few years.

Why can't we increase the main frequency?

A physical constraint that affects the CPU frequency is that the main frequency is inversely proportional to the delay of signal transmission between transistors, that is, the higher the transistor density, the higher the clock frequency, which is why more advanced technology can be adopted to improve the CPU frequency before 2003, and the improvement effect is particularly obvious.

However, the frequency increase of CPU is not infinite. This factor is the problem of energy consumption heating. Excessive energy consumption will lead to excessive CPU heating, which may lead to CPU burn-out, and the energy consumption of CPU is roughly proportional to the cubic of clock frequency, that is, the frequency will double, and the energy consumption may reach 8 times before.

The relationship between the main frequency and power consumption of FX8350 and FX9590 has been calculated before, and the relationship between power consumption and frequency improvement has been roughly verified, because FX9590 is the official overclocking version of FX8350, with the same process architecture and the same number of cores, and the relationship between frequency and power can be well observed. The default frequency of FX8350 is 4Ghz, and that of FX9590 is 4.7Ghz.

The frequency of FX9590 is 1. 175 times that of FX8350, and the cube of 1. 175 is 1.62, that is to say, theoretically, the energy consumption of FX9590 is 62% higher than that of FX8350. Comparing the TDP of the two, we can find that FX9590 is 76% higher than FX8350 (220 divided by 125 and then subtracted by 1). From this result, it is true that the power consumption of CPU is approximately proportional to the cubic of clock frequency. In short, it is certain that the relationship between frequency and energy consumption is not linear.

Of course, some people will say that since the increase in energy consumption leads to fever, the adoption of advanced technology can alleviate the problem. Theoretically, it is possible, but the more advanced the technology, the higher the heat density and the easier it is to accumulate heat. For example, although the 7nm process can provide lower energy consumption than 14nm, the heat accumulation problem of the 7nm processor is more serious. Although the energy consumption is much lower, the temperature will not be higher than that of 14nm products.

A CPU contains billions of transistors. For example, Intel's mainstream CPU has 2 billion transistors, and some high-end products have as many as 6 billion transistors. Transistors will generate dynamic power consumption according to the main frequency of CPU when converting analog signals, so the higher the main frequency of CPU, the greater the calorific value.

Of course, the manufacturing process of chips has been developing continuously. According to Moore's law, the number of components that an integrated circuit can accommodate will double every year and a half, and the performance will also double.

Pentium 4 processor in 2000, manufactured by180 nm; ;

Core i7-980X comes from 20 10, and its manufacturing process is 32nm；.

Core i7 4960 x 2065 438+03, manufacturing process 22nm；;

Now the manufacturing process of Core i7 9700k has reached the level of 10nm. The smaller the transistor, the lower the turn-on voltage, which can compensate for the increase of power consumption caused by the increase of CPU main frequency.

However, the manufacturing process of CPU will not be improved endlessly. The more difficult it is, the biggest obstacle to improve the main frequency of CPU is the manufacturing process. Moreover, the size of transistors is reduced, but the increase in the number will highlight the problem of heat accumulation between transistors, so the total heat generation will not be reduced too much.

Moreover, the main frequency is only one aspect of CPU performance and does not represent the overall performance of CPU. The performance parameters of CPU include secondary cache, tertiary cache, instruction set, front-end bus and so on. Blindly increasing the main frequency of CPU will double the calorific value of CPU. Finally, in order to cool the CPU, we will spend a lot of effort on the heat sink, which is not worth the candle.

Therefore, in order to improve the speed of CPU, semiconductor engineers can design multi-cores for CPU and achieve the same effect. Just like there are 100 arithmetic problems to be solved, a single-core CPU is for a quick calculation expert, and a multi-core CPU is for four people with average quick calculation ability, but in the end it takes four people to complete 100 in a short time. After all, many hands make light work.