Intel's Pentium processor, brings two technical concepts, one is superscalar, which allows the CPU to execute two instructions in a cycle.

The other is the depth of the flow, that is, the flow depth of an instruction is divided into multiple steps, and the flow depth of the Pentium processor has reached 14 levels, and when it has developed to the Pentium 4 processor, the flow depth has reached more than 20 levels.

Super-flow technology has a significant role in improving performance and frequency, but it is not without its flaws.

One of the biggest drawbacks is that the ultra-long flow of water faces inefficiency in execution.

This is also the Ben Si era, the Intel Pentium 4 processor with the same main frequency is often pressed to the ground by AMD Athlon's processor with a similar frequency.

And hyper-threading can solve this problem perfectly.

In this era where the main frequency is paramount, Huaxin Technology's CPU wants to make a name for itself in the market, so it naturally considers the main frequency.

There are only two ways to increase the main frequency, one is to improve the process, and the other is to improve the technology.

In terms of process, Huaxin Technology has no choice, only to find a way in technology.

Wang's solution is to first use a large number of RISC reduced instructions in the instruction set of the CPU, because each instruction of the computer requires additional transistors and electrical components, and most complex instructions can be combined with multiple reduced instructions, so as to choose a good balance between efficiency and performance.

The second is to choose super flow technology, the first generation of Huaxin Technology's flow design is 18 levels, and the design frequency is 200Mhz, which naturally has the defect of insufficient assembly line execution efficiency.

The original design of this CPU, such as technical indicators, floating-point, integer and other calculations, is no weaker than the Pentium processor in the world.

Coupled with hyper-threading technology, Wang Anran is looking forward to the results of the simulation.

But the facts gave him a resounding slap in the face.

EDA simulation test results show that SMT hyper-threading technology does not bring significant performance improvement to the CPU.

Spending 10% of the number of transistors to build only brings about a 10% performance improvement, which is already a chicken rib.

There is no need for Huaxin Technology to explode such an important technology for this 10% performance improvement.

Wang Anran analyzed the reason, it should be that the eighteenth level has not yet explored the potential of ILP flow, and the CPU design of this era is limited by the number of transistors, and its own idle resources are also limited.

The question is, where do you add these 10% transistors?

This is not a problem at all for a master like Wang Anran, there are many ways to improve CPU performance, and the best one is suitable for this era.

The final solution is that Wang Anran will add AMD's mature voltage prediction algorithm to the new CPU.

The goal of branch prediction is to store the branch target in the cache of the branch target in advance when the CPU processes the process and when the branch occurs.

The essence of the branching algorithm is to guess the path of the instruction flow.

So the question is, what if you guess wrong?

The solution to guess wrong is very simple, that is, guess again until you guess correctly.

But the consequences are obvious, the process stalls, the CPU execution is inefficient, and the user's experience is that the computer is stuck.

Therefore, an excellent branch prediction method often has a huge performance improvement in CPU execution efficiency.

......

Fu Yan is a graduate student of the Academy of Sciences, received a doctorate in microelectronics from the University of Cambridge, and spent three years at the Texas Instruments European Research Institute.

This is the turtle talent that Huaxin Technology has spent a lot of money to poach.

Fu Yan has a deep knowledge of computer algorithms and has published articles in many heavyweight journals at home and abroad.

I originally thought that coming to Huaxin Technology would be overkill, but I didn't expect to enter the CPU division as soon as I came in.

Huaxin Technology has such ambitions?

As a Chinese, Fu Yan, like most of you, has a Chinese core dream.

Can Huaxin do a good job in China?

Three days later, Fu Yan turned from doubt to affirmation and amazement.

From the instruction set to the CPU architecture design, Huaxin Technology has an extraordinary background.

Is this still a start-up? From a few dozen people to less than 200 people now, it can support the big task of CPU design?

And Fu Yan can also give full play to his expertise in mathematical algorithms here.

But today he has seen the methods of his legendary founder.

The strategy of changing the prediction table was simply changed, and several prediction strategies were added, and the accuracy of the whole prediction was increased by 30%.

Wang Anran shook his head, this is his learning from other people's achievements, there is nothing to boast about, the optimization of a prediction table strategy, which is the crystallization of the wisdom of tens of thousands of scientific and technological personnel by concentrating the development of dozens of generations of processor technology.

Compared with the original version, the CPU branch prediction algorithm integrated in Huaxin Technology can only be regarded as a castrated version, which is no way, the original algorithm must be fully implemented, and 3 million transistors are not enough.

And Wang Anran's low-key was serious in Fu Yan, and it became an unfathomable meaning, which not only hit his arrogant self-esteem, but also made him calm down and continue to finish the following things.

"Professor Fu, if you want to make branch prediction have better room for development, there are higher requirements for the width of the branch prediction table and the pre-read data capacity, and it is necessary to design a second-level cache under the CPU cache. ”

Fu Yan shook his head and said: "It's good, but it doesn't work technically, the second-level cache needs to allocate additional bandwidth on the bus, which will inevitably affect the communication between the CPU and the memory."

Add to that the fact that SRAM caches are not only expensive, but also take up a lot of chip area, and we won't have enough transistors for other functions. ”

Wang Anran shook his head, the more sophisticated the talent, the easier it is to be limited by conventional thinking.

Intel is the industry hegemon, but it has also committed the crime of "creatively" integrating the CPU L2 cache into the CPU core in order to prevent the huge SRAM cache from occupying a limited CPU core area in the Pentium III era, resulting in the CPU L2 cache can only run at half speed.

As a result, the CPU runs inefficiently and has weak performance, which is inferior to AMD's K6 architecture CPU in the same period.

"Professor Fu, we can design a separate bus communication bus for the L2 cache, but we need about 2% transistors. ”

"Redesign a bus?"

Wang Anran nodded and continued: "There is no need to use SRAM cache for L2 cache, EDRAM cache can also be used, plus 128kb L2 cache, the performance is worth looking forward to." ”

Fu Yan was surprised, the boss's thinking was really whimsical and eclectic, but he still raised his questions.

You must know that all today's chip designs, there is no case of edram as a CPU cache, Huaxin Technology's first CPU, is it necessary to be the first company to eat crabs?

"Mr. Wang, is this really OK with EDRAM?"

Of course, although it has not been tried in this era, in 9102, EDRAM cache is already a very mature technology.

Wang Anran nodded affirmatively, "If you don't try, how do you know if you can't ......?"

............

It is undoubtedly the most important step to take the right step to engage in EDA design software at the beginning, and with the help of EDA software, many difficult problems have become simpler.

For example, with EDA software, the designer does not need to know how the transistors are logically processed at the bottom layer, he only needs to compose the existing logic gates into logic units, and then design these cells into larger computing units.

This undoubtedly greatly reduces the threshold of design and the difficulty of design.

The role of the underlying engineer is to look for simpler transistors and electrical components in the face of these abstract logic gates and logic units, and to display them on a physical level.

It's like a pipeline, it was originally a workshop-style design, and now there are two pipelines, one is to design the CPU, and the other is to design the circuit.

In 9102, this assembly line will be deepened into dozens of design categories, and the specialized categories of EDA software will also be improved.

Huaxin Technology is now developing in this direction.

The redesigned CPU performed well in EDA simulation tests, delivering a 37% improvement over 200MHz plus pure hyper-threading technology.

The number of transistors has increased by 16%.

Acceptable range.

Professor Fu pointed to the test results and said: "Mr. Wang, I really didn't expect that with the addition of L2 cache and branch prediction, the CPU performance would be so improved, with a nearly 50% performance advantage at the same frequency. ”

Wang Anran said: "It's a pity that I can't go up to 200MHz." ”

CPU frequency is affected by the bus and L2 cache, the heat will not be controllable at the frequency of 200mhz, but at 166Mhz, plus good heat dissipation should be possible, which depends on the specific test situation after the tape-out is produced.

But even so, compared with the latest color Pentium 586 in the world, the frequency of 133Mhz is still very surpassed.

It seems that it is completely possible to press Intel to the ground and rub it......

............

A month later, Wang Anran returned to the realization of the public, and Wang Anran had already made a decision about Huaxin Technology's CPU plan.

Clark contacted Wang Anran as soon as possible, and in Wang Anran's office, Clark expressed his concern.

"Mr. Wang, can you tell me clearly that Huaxin Technology has decided to enter the CPU field. ”

Wang Anran nodded.

Clark sighed and said, "What a disappointing news, is Mr. Wang thinking of using investors' money to test unfamiliar projects? ”

Wang Anran said with great interest: "Mr. Clark, why do you say that?"

Clark doesn't even want to explain, a three-year-old child can see the truth, you can always not understand!

If CPUs are really so good, why are there only a few companies in the world, such as Intel, AMD, Milinda, and IBM?

"Mr. Wang, if your company insists on entering the CPU industry, Goldman Sachs may be skeptical about your company's future development, which will have a significant impact on the valuation of Huaxin Technology. ”

Wang Anran took a sip of water and said calmly: "Mr. Clark, what if Huaxin Technology succeeds!"

()