Past events: The founder reveals how Bitmain S9 became the "king of machines"

Preface: With the collective entry of high-computing mining machines, the Antminer S9, a first-generation king of mining machines, is about to exit the stage of history. Looking back on history, the author believes that the Chengdu team achieved a cross-generational advantage over competitors through dynamic triggers, which is also the main reason for the success of S9. The design of S9 is a miracle of Bitcoin mining machines. It has become the most produced mining machine among Bitcoin mining machines. With a life cycle of more than three years, it is called the king of mining machines.

(The author of this article is Xie Dan, former technical director of Bitmain, and Wu Says Blockchain authorized the publication. The views in this article are the author’s personal opinions and do not represent the position of Wu Says Blockchain)

In August 2014, I opened a backend design service company in Chengdu. I wanted to provide value-added backend design services. When looking for clients, I searched online and thought that companies that make Bitcoin mining machines should have this demand, so I wrote an email to the mailbox on the Bitmain website at the time:

At that time, I wrote emails to many integrated circuit design companies, and Bitmain was the fastest to respond. Later, when we chatted, I learned that this email was used by Mr. Wu Jihan at the time, and Bitmain was also looking for a solution to improve the competitiveness of its chips. So Wu Jihan forwarded this email to Mr. Zhan Ketuan, who met me when he passed by Chengdu in September. Mr. Zhan and I met twice, and we had a good chat. Mr. Zhan felt that this project would take a long time, was risky, and had concerns about intellectual property rights, so he suggested that I dissolve the company and lead the team to join Bitmain.

In October 2014, I joined Bitmain with a small team of two people and became the full customization department of Bitmain. The initial direction was the domino logic mentioned in the email.

Domino logic is a relatively mature dynamic logic structure, and its main form is as follows:

Compared with the corresponding static circuit, the domino logic circuit has the following advantages and disadvantages:

1. Because the signal only needs to drive the N-tube, the capacitance is smaller and the speed is faster.

2. Because point A is a dynamic point, the final inverter M3/M4 is necessary.

3. There is a charge-sharing problem at point A.

4. M1, M2, M3, M4 are newly added devices, and PMOS is a reduced device

5. Because M1 and M2 are driven by clock signals, their power consumption is at least twice that of other common signals.

Therefore, considering the power consumption, M1 and M2 are clock signals, so their power consumption is at least twice that of the signal. The power consumption of M1+M2+M3+M4 is the power consumption of 6 MOS tubes. For general logic, there are not many standard cell libraries with more than 12 MOS tubes. In terms of area, M2+M4 increases the area of ​​2 NMOS tubes, and although the number of PMOS is reduced, it is difficult to reflect the saved area on the layout. In other words: although domino logic will speed up the chip, the area and power consumption will increase slightly.

When we submitted our conclusions in January 2015, we basically confirmed that Domino logic is not suitable for mining chips. Bitcoin mining chips are purely parallel computing, so speed is not that important. The biggest cost of mining machines is electricity, so power consumption is the most important. When we evaluate the area and power consumption rule of thumb, the weight of power consumption is more than 3 times that of area and speed.

The failure of our domino logic experiment did not discourage us from exploring, because we found a standard cell that is suitable for dynamic logic with a large area and many transistors, and an obvious example of this standard cell is the flip-flop. So we went back to the beginning and took the dynamic flip-flop as our target.

At the beginning of integrated circuits, in the 1970s, because the cost of each transistor was high, the triggers at that time were all dynamic, whether it was domino flip-flop, C2MOS edge-triggered flip-flop, or TSPC positive-edge flip-flop, etc. We discovered a treasure trove from it.

For example, a TSPC positive-edge Flip-flop has the following logic:

If we still count the clock gate as two gates, this flip-flop is equivalent to 4*2+7=15 gates in total.

The structure of the static trigger we used most often before is as follows:

Adding the inverter of clk, we have 8*2+12+4 =32 gate equivalents. The power consumption of the equivalent transistors of the static logic trigger is more than 1 times that of the dynamic logic.

Similarly, in terms of area, dynamic logic has 11 devices, while static logic has 22 devices, which is exactly twice the area.

After we confirmed the use of dynamic logic triggers, the next step was how to integrate it into our design process. We finally added some time constraints to the functional description of the static logic trigger to prevent leakage current in the dynamic logic from causing leakage at the dynamic capacitor point. And when the timing and power consumption are improved, some parameters of the static trigger are applied. Simply put, we added an outer frame to the dynamic logic so that it looks like a normal static trigger to the front-end designers, and there is no difference for the front-end design and synthesis.

After completing the dynamic trigger library, we have a new trigger with half the area and half the power consumption and almost the same timing. Because Bitcoin needs to calculate continuously, it is unnecessary for static triggers to save data for a long time. After we complete the dynamic trigger, Bitcoin's distributed computing and fully pipelined logic are particularly suitable for dynamic triggers. For the pipeline, its structure is as follows:

We can directly change it to:

Considering that the area and power consumption of the new dynamic trigger are only half of the original ones, the new pipeline has doubled the speed with the same area and power consumption. In our Bitcoin mining chip, we changed the pipeline from 32 to 64, doubling the computing power.

We completed the design of the 28nm BM1385 chip (Antminer S7) in mid-2015, and completed the design of the 16nm BM1387 chip (Antminer S9) at the end of 2015. In terms of performance, our 28nm is almost the same as the 16nm of our competitors, and our 16nm chip is half the cost of our competitors. Through dynamic triggers, we have achieved a cross-generational advantage over our competitors. In particular, the design of S9 is a miracle of Bitcoin mining machines, becoming the most produced mining machine among Bitcoin mining machines, with a life cycle of more than three years, and is known as the king of machines.

In the digital currency mining industry, which requires almost no software environment, a product costs only half of the competitor's, which is a huge competitive advantage. This means that you can have the advantage of price wars as you wish. Because you can't make money by selling to your competitors, but Bitmain still has a gross profit margin of more than 50%. It is precisely by relying on the secret weapon of dynamic logic and the successful sales of S7 and S9 that Bitmain has gone from a situation of less than 20% market share to a single dominant player (more than 70% share).

A direct result of Bitmain’s rise is the exit of foreign Bitcoin chip companies from the market. In 2014 and 2015, high-profile companies such as KNC, Bitfury, Spondoolies-Tech, and 21 Inc. all quickly declared bankruptcy or withdrew from the mining chip market.

With the departure of some Bitmain employees, the dynamic trigger technology has gradually spread to other domestic chip developers, but this technology is still basically restricted to China. In 2017, Japan's GMO was still trying to enter this field at 12nm and 7nm. Judging from the company's publicity, they still adopted the static trigger structure. Coupled with the bear market cycle in 2018, it is foreseeable that they will exit the market at a loss one year later.

Before 16nm, the cost of a new generation of process masks was within a few million dollars, while the technology, manpower and risk of redesigning dynamic logic were obviously higher than millions of dollars, so the advantages of dynamic logic could not be reflected. However, after 16nm, the cost of the new 10/7nm and the future 5nm are tens of millions of dollars. Dynamic logic can surpass the performance of a process node, which will make it more advantageous and rejuvenate. I look forward to the emergence of more companies that apply dynamic logic, so that this ancient design art can be revived in our era.