ETH, ZEC professional mining machine disassembly evaluation & future direction of mining hardware

Once upon a time, one of the slogans of the currencies using the Ethash algorithm represented by ETH and the currencies using the Equihash algorithm represented by Zcash was ANTI-ASIC, which means that they have a certain ability to resist the emergence of ASIC chips. In essence, the principle of the anti-ASIC ability of these two algorithms is to use memory-demanding algorithms, that is, a certain amount of RAM storage is required during calculation. The reason why the logic of "using memory-demanding algorithms = anti-ASIC" was established in 2014 and 2015 was closely related to the hardware manufacturing cost at that time. However, as time went by, the price of RAM dropped, and the manufacturing cost of dedicated mining machines with a large amount of RAM also dropped, so there is today's situation - various algorithms have been conquered by mining machine chip manufacturers, and dedicated mining machines for the corresponding algorithms have been listed.

Today, I will disassemble the hardware of Bitmain E3 and Z9 to give you some scientific knowledge: What are the differences between ASIC chips of these algorithms and ASIC chips of old algorithms such as Bitcoin and Litecoin?

Ethash (ETH) algorithm-specific mining machine Bitmain E3 disassembly review

Background knowledge:

The Ethash algorithm needs to build a DAG file in memory, and frequently read the contents in memory during calculation to obtain anti-ASIC properties in disguise. Currently, this DAG file is close to 3G, which means that only graphics cards with more than 3G video memory can be used to mine ETH. Generally, the higher the memory bandwidth, the higher the computing power of the algorithm.

E3 parameters:

The hashrate is 190MH, the power consumption is 760W, and there are 18 hashrate chips. The efficiency is 1.5 times that of a graphics card mining machine.

Rumors of ETH's professional mining machine appeared in March this year, and Bitmain officially released the E3 mining machine in April. It was not until recently that the first batch of E3 mining machines arrived, and we could really take a peek at the internal structure of this mining machine.

(The following pictures are taken from a YouTube video, posted by Justin A)

After opening the outer shell, we can see three main computing boards compactly packed inside the mining machine.

E3 structure after opening the shell

Pull out a hashboard, and you can see that each main hashboard is connected to 6 sub-hashboards. The computing chip in the middle of each sub-hashboard is covered with a heat sink, surrounded by a circle of memory chips, and its appearance is similar to that of a discrete graphics card used in a laptop.

Main hash board structure

After removing the sub-computing board and the heat sink, the true face of the ASIC chip can be seen: the BM1790 computing chip. The memory chip on the side is DDR3 128MB, with a total of 4*8=32 chips. The memory capacity of a single chip is 4GB, and the memory capacity of the entire mining machine reaches 72GB. The chip uses the product of Taiwan ESMT. Although the memory selected by E3 is currently a relatively cheap DDR3 solution, the usage of 72GB of the entire machine makes its manufacturing cost not low.

Sub-hash board structure

After reading the disassembly diagram, everyone has a sufficient understanding of the structure of the current professional mining machine for the Ethash algorithm. Since the computing power is proportional to the memory bandwidth in the Ethash algorithm, and the threshold of a single computing chip must be matched with 4GB of memory, the possibility of the appearance of the previously rumored Ethereum super mining machine that can fight one against a hundred is zero. But at present, in addition to E3, there are ETH professional mining machines from other manufacturers on the market, and the performance parameters are even more exaggerated. For example, the product A10 launched by another domestic manufacturer, Xindong, has achieved an exaggerated computing power of 485MH at a power consumption of 650W. Damao speculates that A10 uses a more advanced memory chip with a larger bandwidth (DDR5) to achieve the purpose of increasing computing power, but such an operation will undoubtedly make the manufacturing cost very high.

Equihash (ZEC) algorithm dedicated mining machine Bitmain Z9 chip level disassembly evaluation

Background knowledge:

The algorithm used by ZEC, ZenCash, ZCL, etc. is actually Equihash (200,9). BTG used to be the same as them, but after being attacked by 51% in May, it has now changed to Equihash (144,5).

The Equihash algorithm has parameters that affect the solution, called the N and K parameters. They affect how long it takes to solve and how big the solution will be for different combinations of processing power and memory. The parameters chosen by Zcash before its launch were N=200 and K=9. After Zcash was launched, cryptographers determined that the values ​​144 and 5 seemed to produce smaller, more ASIC-resistant solutions.

In this case, ASIC resistance means requiring more memory, with parameters 200, 9 having a minimum memory requirement of about 140MB, while parameters 144 and 5 require nearly 2.5GB of memory. Integrating memory into ASIC miners is expensive, so hashing algorithms with higher memory requirements are more ASIC resistant. We can refer to this version of the Equihash algorithm as Equihash (144,5) to distinguish it.

Z9 parameters:

The computing power is 40.8Ksol and the power consumption is 1150W. The efficiency is about 40 times that of the graphics card mining machine.

SRAM:

Static random access memory is a type of memory that is often designed inside a chip. For example, the L1, L2, and L3 caches of a CPU. Increasing SRAM will significantly increase the chip area and increase chip manufacturing costs.

(The following pictures are from the bitcoingold forum, author MentalNomad)

No memory chips were found on the Z9’s hashboard, but after removing the heat sink, it can be seen that the core area of ​​the BM1740 chip used by the Z9 is particularly large, and it is speculated that the Z9’s RAM is integrated into the chip.

Z9 computing chip

After destructively removing the chip and peeling off the top layer, the basic structure inside the chip can be seen.

Put it under a microscope to enlarge it and see it clearly

As you can see, the chip has 144MB of memory integrated inside, which just meets the requirements of the Equihash (200,9) algorithm. As for why the storage is integrated into the chip at a high cost instead of being external, Da Mao guessed that it was for the purpose of increasing computing power. After all, the read and write speed of the internal cache of the chip is N times that of ordinary memory chips. However, it is precisely because of this that the 2.5G memory required by the Equihash (144,5) algorithm does have the ability to resist ASIC. After all, it is too difficult to meet the large capacity of 2.5G and the ultra-high read and write speed to ensure the leading computing power, and the manufacturing cost cannot be too high.

Through the disassembly and evaluation of the above two machines, Da Mao also summarized:

1. Although the two mining machines use different solutions to add memory to meet the needs of the algorithm, it is enough to show that the use of memory-demanding algorithms cannot completely prevent the emergence of professional mining machines, but can only delay the emergence of professional chips.

2. When specialized chips for memory-demanding algorithms emerge, they will have a lead in computing power compared to GPUs. However, compared with the fact that the Bitcoin SHA256 algorithm ASIC was hundreds or thousands of times ahead of GPU computing power, the lead is not too exaggerated.

FPGA returns to the market, is the competition in mining equipment entering a new era?

Driven by the need to strike a balance between computing power, manufacturing cost and flexibility, FPGA chips, which once dominated the market in 2013, are showing signs of making a comeback.

FPGA, short for gate programmable circuit, is a semi-customized circuit in the field of application-specific integrated circuit (ASIC), which not only solves the shortcomings of customized circuits, but also overcomes the shortcomings of the limited number of gate circuits of the original programmable devices. FPGA is generally slower than ASIC (application-specific integrated circuit) and has a larger circuit area than ASIC to achieve the same function. However, they also have many advantages such as fast production, can be modified to correct errors in the program and cheaper cost.

FPGA can be used for mining after proper hardware programming, and its efficiency is not low. However, its disadvantages are that it is difficult to deploy in large quantities and the hardware programming threshold is high.

After Monero chose to change its algorithm and hard fork due to the emergence of dedicated mining machines this year, mining machine manufacturers did not dare to continue to develop new chips for the new algorithm directly, because the new chips would likely be boycotted by the community again and face a situation where they would be unavailable. Under such circumstances, FPGA has found its space for survival. Da Mao has found some hardware products developed by foreign enthusiasts, such as the following one:

The introduction of FPGA mining machines with low barriers to entry, suitable for ordinary people to use, and capable of providing new algorithm upgrade services may be the next direction of research and development for mining machine manufacturers. As long as POW survives, the game between algorithm researchers seeking a new algorithm compatible with every hardware product and hardware developers seeking a particularly efficient mining hardware will always exist.