This content comes from @What is worth buying APP, and the views only represent the author's own |Author: iN in
Today we're going to talk about something that a lot of digital bloggers won't tell you at all.
The origin of this matter lies in this reply:
Can you ask, "Are there any other instruction sets that are currently affecting CPU performance?" In this way, it actually shows that the reader is starting to pay attention.
CPU is a fairly complex product, and its complexity is beyond the imagination of the average person. Faced with choice, it is also difficult for many digital bloggers to understand the advantages and disadvantages of different technologies used by different CPUs for ordinary users. Faced with such a problem, everyone abstracts the originally complex product parameters into general "technical representations" such as how many frequencies, a few nanometer processes, several generations of CPUs and several cores.
Although this kind of cognition is basically sufficient in the face of daily life, it is very rudimentary. Why do you say that, let's first talk about the plot of a novel that iN recently read:
The Xiongnu's Mao Dun Shan Yu tangled 300,000 cavalry to attack the land of the Han people, and at the oath meeting, Mao Dun shouted, "Men of the grassland, we must step on the Great Wall, grab the land in the south, let the Han people herd sheep for us, and turn the fields into pastures"......
As Han Chinese, do we feel that something is wrong? "Why do you only know how to graze?" This is the problem of "cognitive limitations". As the northern nomads at that time, it was difficult for them to imagine that Yingche Jiasui was full of traffic; It's harder to imagine that the soft red is drunk with gold.
The complexity and breadth of the technology landscape is also far beyond the comprehension of the average user, especially when it comes to the details of CPU technology. Ordinary users, like Maudun, only know that the land can be grazed.
Not convinced? So let's ask a few questions about soul torture:
Why are Intel's Core CPUs divided into i9, i7, i5, and i3 product lines? What exactly is the difference between them?
Why did CPUs start to divide into large and small cores starting from the 12th generation Core?
Why did the iPad Pro, which was released two days ago, use the M4 CPU before Apple's notebooks and desktops?
Why are Intel and AMD CPU packages and cores different?
Why are CPUs divided into consumer and professional markets?
……
If you don't know the right answers to these questions, then in the digital realm, you're a "sloppy" who can grab land but doesn't make good use of it, and you might be complacent about buying the wrong CPU.
Let's talk about the deeper reasons why Apple uses the latest M4 CPU on the iPad. This is the topic that iN's friends have discussed the most with iN recently. A lot of people think that Apple is too aggressive to develop a product that performs against the sky. The deeper reason that iN believes is that Apple's M4 chip production line has insufficient capacity and the yield rate is stretched.
How do you say that? You need to understand the CPU generation and manufacturing routines.
APPLE'S CPU IS GENERALLY DIVIDED INTO FOUR DIFFERENT GRADES, AND ON TOP OF THE BASE CPU, THERE ARE THREE HIGHER LEVELS OF PRO, MAX AND ULTRA.
So what would you do if you were to open a production line to produce Apple's CPUs? Is there a separate production line for each different grade of CPU? Isn't that too costly?
In fact, any manufacturer of CPUs will etch the most complex and high-end CPU products on the wafer.
In the manufacturing process, if the manufactured chip is defective, some cores on the chip will be shielded, and the chip will be downgraded for sale.
APPLE IS EVEN MORE EXTREME, THE AREA CONSUMED TO MANUFACTURE AN M ULTRA CPU ON THE SILICON WAFER IS 8 TIMES THAT OF AN ORDINARY M CHIP, WHEN THE POSITION OF THE 8 M CHIPS IS ALL QUALIFIED, YOU CAN PRODUCE AN ULTRA, OTHERWISE THE WAFER ON THE AREA OF THE 8 M CHIP WILL BE CUT, AND IT WILL BE CUT INTO A MAX CHIP AND A PRO CHIP, AS WELL AS TWO M CHIPS, OF COURSE, THERE IS A HIGH PROBABILITY THAT ONLY ONE OF THE TWO M CHIPS CUT OUT CAN BE USED,
This is because of the compromises that have to be made on a wafer due to yield issues. Not only that, but some of the abolished individual dies (bare die) themselves are not completely bad, and they can continue to be used at a lower grade after shielding some substandard circuits. For example, when we choose an Apple computer, we have an option for CPU selection:
Allows users to choose between a 10-core or 12-core processor, and whether to use a 16-core graphics processor or a 19-core graphics processor at the time of purchase. If this option is reversed, it reads:
It's all written with 12-core processors, but there are a batch of defective products that have broken two cores, and there are all 19-core graphics processors, but this batch of defective products has broken 3 graphics cores, and if you are willing to buy it, it will give you a price reduction of 2250 yuan.
Therefore, the CPU with the most defects and the worst performance is the defective product that occurs when producing the highest-end CPU. These CPUs will quickly sink into the next level of the product.
For example, this time the Ipad Pro
You'll even find that iPads of different capacities come with different CPU cores.
Therefore, in Apple, when an ULTRA is produced, the number of cores will be divided according to the integrity of the core, and it will be cut into MAX, PRO and basic chips according to the quality of the bare die, so that the differentiation of the product is made.
Similarly, in Intel and AMD, CPUs of different product lines are divided according to the quality of the bare die, such as i9, i7, i5, i3, and the frequency index of the CPU is reduced from the highest frequency to a specific frequency to avoid the defects that already exist on the bare die. In this way, you should understand why the Core CPU is divided into i9, i7, i5, i3 different product lines, and there are so many frequencies in different product lines, right?
For large and small cores, it's a different topic - instructions
Since the 12th generation of processors, Intel has begun the design of hybrid architecture, which is commonly known as large and small cores. It's really a process of deredundancy.
CPU cores are divided into performance cores (large cores) and performance cores (small cores). In terms of microarchitecture, the performance core adopts the Golden Cove microarchitecture:
The performance core uses the Gracemont microarchitecture
From the architecture diagram, we can see the differences between the two cores, but at the same time, we can also see the two cores and their similar functions. These very similar functions are actually the main functions of the CPU.
The small core does not have the performance of the large core because it is small, but it does not make a circuit that can complete the instructions that the large core can complete.
For example, the AVX-512 instruction set can run all functions on large cores, but only some functions can be supported on small cores.
Therefore, if the AVX-512 function is enabled while running the game, the load on the large cores will increase immediately, but the load on the small cores will not change at all.
Another example is FP16 (Half-Precision Floating Point), that is, 16-bit half-precision floating-point calculations can only run on large cores, and there is no support on small cores. There are many similar commands, such as AMX, VNNI, TSX......
So say – what is a directive?
From the perspective of low-level programming, CPU instruction refers to the type of operation and operation mode that the CPU can perform defined by opcodes and operands
However, from the perspective of the hardware design of the CPU, an instruction refers to a circuit with a specific function and a series of electrical signals that activate this circuit.
So there is a very important concept - if you want more functions, you need more chip area and more power consumption, which is ultimately higher procurement and use costs. The cheapness of having the horses run and the horses not grazing exists only in the imagination. There is no such thing as a cheap, fast, and low-power CPU for CPUs.
Similarly, due to the different circuit configurations inside various CPUs and the corresponding market demands, CPUs are divided into consumer CPUs and professional CPUs.
Because of this, iN usually tells friends that people who want high performance with industrial CPUs like J1900 are either stupid or stupid. That's not how cheap things are used.
Back to today's topic, how to choose a CPU: a very simple judgment, buy the most expensive CPU you can afford on the premise of a fair price. This is determined by the manufacturing principle of the CPU, only the highest-end CPU can be closest to the original design index of the original design of this series of CPUs, in each generation of CPUs, except for the top-of-the-line model, all CPUs are actually more or less castrated versions.
This kind of castration is not determined by the assembly process of the product, and it is not like when we buy a car, unnecessary configurations can be reduced, but only through shielding to hide the defects of the CPU in the manufacturing process. The lower the CPU purchased, the more defects there are in the CPU itself. This is also the reason why a question that can explain many people:
Many people are discussing why the power consumption of the same model of CPU is different, and even derives the difference between CPU constitution, read the above article, you should understand - in the manufacturing process defects are random and diverse, so even the same model of CPU due to the "broken" part of the difference will bring about a difference in "constitution". The model number of the CPU sold is just a range identifier to determine the degree of CPU damage.
In addition, many people have learned to ask if this graphics card supports "ray tracing" when buying a graphics card. In fact, on the graphics card, when the "ray tracing" finally sinks to the circuit, it is a set of circuits, that is, a set of instructions. However, few people ask whether the purchased CPU supports AVX-512, whether it supports AV1 hardware decoding, and whether it supports AES acceleration...... The reason is that these instructions of the CPU are not sufficient for the average user. But if we want to get a CPU parameter menu, such as the following one:
Most users simply don't know how much of an impact these dense parameters can have on their system experience. For example, when some functions are not enabled, the system will run like this:
Note that the CPU usage is fairly flat and smooth.
But if you turn off some of the settings, you'll notice that the CPU usage spikes right away:
However, in the feeling of ordinary users, it may not be clear why the speed is slow, and the accumulation of this phenomenon leads to the user's feeling that the computer is "one card and one card". This is one of the main reasons why many users feel that their computers are slow.
Sometimes upgrade to a new CPU, this CPU feature provides direct acceleration of certain functions, and users will feel that the stuttering phenomenon is alleviated. More often than not, a CPU provides some features but reduces them, making the user feel that the system has not been significantly faster than the CPU has changed.
This phenomenon is actually due to the fact that the user will not be able to read the CPU feature table, and whether the skills will match their common tasks are missing.
At the same time, there is another important point – this is a watershed period. Many AI applications are gradually moving from the cloud to on-premise. In recent months, iN's buying advice to friends has really been about choosing a processor with an NPU.
The reason is that the adaptive application of NPU on personal computers is still not very popular, but under some special tasks, NPU itself can already highlight its existence under a suitable framework.
Especially on some small machines, the value of NPU for AI is much higher than the value of a mid-range graphics card.
The author declares that there is no interest in this article, and friends are welcome to communicate rationally and discuss harmoniously~