Why I believe FPGA mining machines will erode the computing power share of GPUs before the emergence of ASICs

Author: Yan Xin, works at blockchain investment institution Fundamental Labs

With the launch of the new popular PoW public chain mainnet such as Nervos, the process of PoW public chain computing power competition and balance before the emergence of ASIC mining machines has once again attracted attention. Generally speaking, the degree of matching between the mining algorithm and the characteristics of the mining machine device affects the market share of a certain type of mining machine, and during the launch of the Nervos mainnet, the performance of FPGA mining machines is quite conspicuous.

Although FPGA mining machines, which are often mentioned and discussed in the history of the mining industry, have never been mass-produced and sold, their impact on the ecology of the PoW public chain cannot be ignored. In my opinion, FPGA mining machines are very likely to erode the computing power share of GPUs before the emergence of ASICs on a large scale, becoming the most powerful computing power tool.

01
What is FPGA?

Let us start by simply popularizing the basic concepts of what is "FPGA" and what is "ASIC".

FPGA is the abbreviation of "Field Programmable Gate Array", which is a semi-custom circuit.

ASIC is the abbreviation of "Application Specific Integrated Circuit", which is the optimal solution for a fixed algorithm. However, ASIC production is expensive, with the cost of one wafer being around 10 million RMB. Its cost advantage can only be reflected in mature, stable, and high-demand application scenarios.

For scenarios such as cryptocurrency mining or machine learning where the computing logic needs to change frequently or the chip demand is relatively small, FPGA has a place with its computing speed comparable to ASIC and the feature of algorithm updateability. FPGA is also often used to verify the computing logic in the ASIC being designed.

In a sense, FPGA is a substitute before the emergence of ASIC, and plays a unique role in scenarios with high uncertainty.

FPGA Internal Structure

FPGAs and GPUs can also be programmed to perform a variety of different computing tasks, but the computing logic of FPGAs is implemented through a logic gate array based on a look-up table (Look-Up-Table), which does not rely on the von Neumann structure. The result of one operation is directly fed into the input of the next operation without having to be temporarily saved in the main memory. Therefore, not only is the bandwidth requirement for the internal memory much lower than that of the GPU, but it also has the characteristics of fast response and lower power consumption.

Therefore, FPGA can show unparalleled advantages over GPU for some algorithms that require less memory and can be processed in pipelines.

02
From “ASIC Resistant” to “ASIC Friendly”

If we want to discuss the future trends of FPGA mining machines in more depth, let us start with the entire process and context of the development of mining machines.

After Bitcoin was monopolized by various dedicated computing devices, a number of early "Altcoins" such as Litecoin began to embody the idea of ​​"Anti-ASIC" (preventing ASIC mining) in the design of mining algorithms. This design is conducive to allowing more enthusiasts to participate in the mining of new cryptocurrencies with ordinary computers, thereby promoting the decentralization of new cryptocurrencies.

However, this kind of "Anti-ASIC" has always been proven to be unsustainable in the long run, and the delicate balance can easily collapse after a certain cryptocurrency boom. Driven by interests, dedicated mining machines corresponding to Anti-ASIC were soon launched.

Although Anti-ASIC itself is indeed an unsustainable process, this gradual failure process has benefited the development of many cryptocurrencies. Therefore, it has been imitated for quite a long time and various Anti-ASIC algorithms have been designed.

Generally speaking, there are three main design ideas for Anti-ASIC algorithms:

• Memory-dependent algorithms, such as Cuckoo Cycle, Equihash, and Scrypt. These algorithms increase the size of the computational input, requiring FPGAs and ASICs to add additional memory outside of the existing computational units, increasing the difficulty of FPGA and ASIC design and reducing their relative advantages.

• Multiple algorithms are mixed, such as X11, X13, and X16R. These algorithms increase the complexity of FPGA and ASIC by introducing several hash algorithms in series at the same time. Later, multiple algorithms in random order were derived, further increasing the design threshold of FPGA and ASIC.

• Regularly change the algorithm. Some cryptocurrencies regularly change the mining algorithm or algorithm parameters, so that dedicated mining equipment is regularly eliminated, which is considered a fundamental anti-ASIC method.

The development of Anti-ASIC algorithms has given GPU mining a lot of room, and has gradually made it the main Altcoin mining equipment, but its side effects have also gradually emerged.

First, GPU miners have no loyalty. The only criterion for most GPU miners to choose mining targets is the daily output. Once the price of a certain cryptocurrency drops sharply, the computing power will also drop sharply. For proof-of-work blockchains that rely on computing power to maintain network security, the robustness of the network will drop rapidly after the computing power drops, making it extremely vulnerable to 51% attacks.

Second, it cannot provide excess returns to its miners. Due to the huge stock of GPUs, the mining revenue of any cryptocurrency that exceeds Ethereum will be quickly flattened by the influx of computing power, so miners cannot get the corresponding incentives to become active builders of the currency.

GPU mining targets are sorted from high to low by daily revenue
ETH ranks third, with a small gap to the top two

Third, the maturity of GPU mining has made the original purpose of Anti-ASIC ineffective, that is, it is impossible to enable more ordinary enthusiasts to participate in mining, and mining is completely monopolized by GPU miners.

Fourth, the stock of GPU itself is huge. In addition to the GPU mining machines that have been put into mining, there are still a large number of GPUs in the hands of MSPs. Under the national strategic background, the National Development and Reform Commission has deleted virtual currency mining and other contents from the eliminated industries in the "Guidelines for Industrial Structure Adjustment". In the future, the entry of real GPU computing power giants is not far away. The excess income of GPU mining relative to the social average investment return rate will be infinitely diluted in the process of the surge in GPU computing power.

In addition to the above four points, regular forks or algorithm changes will lead to the division of the community, and in the long run, it will be difficult for the community to grow.

More and more developers are more conservative about Anti-ASIC, or design rules to guide the hashrate to gradually transition to ASIC in a regular manner. Grin is a good example, but because the Cuckoo Cycle algorithm itself still has very high memory requirements, mining machine manufacturers have not yet successfully launched products as planned.

Nervos, which was recently launched, designed an ASIC-friendly SHA3-based Eaglesong algorithm. This algorithm does not require large memory or involve complex algorithms. Its main content is multiple iterations of a simple hash algorithm, which facilitates pipeline processing in FPGAs and ASICs.

The core of Eaglesong Hash is 42 iterations of the transformation function f

In summary, the transition from ASIC Resistant to ASIC Friendly is entirely possible to be the beginning of a new direction, and FPGA will replace GPU in this direction and become the best mining tool before the emergence of ASIC.

03
Why is it always difficult to produce Altcoin ASIC mining machines?

Since the beginning of this year, mainstream mining machine companies such as Bitmain, Canaan Creative and MicroBT have stopped horizontally expanding their Altcoin ASIC mining machine product lines. The simplest explanation is that the risk-return imbalance is disproportionate.

Compared with GPU, ASIC mining energy efficiency ratio has an absolute advantage of 10 times. Due to the diminishing marginal cost effect, ASIC producers will inevitably tend to overproduce ASIC to grab the computing power share, increase the difficulty, and quickly squeeze out GPU miners. Mining machines for small currencies have small capacity and limited production and sales, which inevitably cause centralization problems.

FPGA is more like GPU here, with a fundamental advantage, that is, FPGA itself is produced by major manufacturers such as Xilinx and Intel. Each FPGA mining machine is nothing more than purchasing FPGAs that have been produced in different places for circuit board-level hardware design and software algorithm optimization. Both of these areas are not as closed and "high-tech" as ASIC layouts, and there will not be much difference between the products of each company, which can better promote decentralization.

According to MRFR statistics, among the major FPGA manufacturers,
Xilinx and Intel have a combined market share of 72%

ASIC is a winner-takes-all game. ASICs designed for the same algorithm using different foundries, different processes, or different design methods may have several times the performance difference. Due to the adaptive adjustment of the difficulty of the blockchain network, computing devices with inferior performance may face brutal elimination. Therefore, each manufacturer strictly keeps the confidentiality of product parameters, which further increases the risk of capital investment.

As a programmable device, FPGA is still widely used in scenarios such as blasting where durability does not need to be tested, even if it is a second-hand FPGA that has been eliminated. It is not worthless like an eliminated ASIC.

Having an ASIC means having a say in computing power, but it does not mean having a say in community governance. If an ASIC harms the interests of most community members, the currency is likely to completely exclude the ASIC designer by modifying the mining algorithm through a fork. For example, in early October 2019, Ravencoin upgraded its mining algorithm to an upgraded version of X16R, "X16Rv2", to prevent ASIC miners on the Ravencoin network from continuing to mine. This algorithm change left many ASIC teams at a loss, and FPGA miners, as the most flexible mining machines for algorithm changes, quickly flooded into the computing power market.

04
FPGA is not a replacement for GPU but a new choice that is more suitable for ASIC-friendly algorithms

In essence, mining is a process of increasing the network computing power through block rewards, thereby increasing the cost of attack. A beneficial mining algorithm should be one that can motivate miners to provide services to the network stably. In this regard, ASIC plays an irreplaceable role, but before the volume of a certain cryptocurrency can no longer support ASIC mining machines, FPGA and GPU essentially play the same role. The difference between the two is the difference in inventory and the difference in the matching degree between their own characteristics and the algorithm design concept.

At present, the availability of GPU computing power (huge stock, stable supply, output and smooth distribution channels) far exceeds that of FPGA computing power. However, if the trend from ASIC Resistant to ASIC Friendly is established and continues to develop, FPGA will inevitably tip the balance of liquid computing power and become a powerful tool for new currency computing power.