To dig deep into the cost of ASIC chip design, we asked the ProgPow core development team nine questions

This article comes fromMediumsecondary title

Odaily Translator |

overview

overview

As long as it is related to the ProgPow and Ethash algorithms, there will be various speculations on the hardware design and development costs of mining machines in the market, usually followed by an authoritative statement: please believe the author who published the prediction, because he/she is in the relevant industry Has extensive experience in the field. Sometimes these speculations are related to the production of cryptocurrency ASIC chips, and other times they are about integrated circuit design.

For those readers who are more familiar with the code, but don't know much about fan-out and rise-times, this article may help them understand the ProgPow algorithm in depth.

(Odaily Jun o-daily Note: Ethash is currently Ethereum’s mining consensus algorithm based on proof of work, and ProgPow is a mining algorithm that attempts to weaken the advantages of ASIC mining machines. Fan-out is a digital signal that defines what a single logic gate can drive Enter the largest amount of jargon. Most TTL logic gates can feed 10 other digital gates or drivers, so a typical TTL logic gate has 10 fanout signals; rise time is a term in pulse technology, The time interval between two moments of voltage rise is the rise time of the network transformer.)

Programmers always give people a sense of omnipotence, from writing scripts to developing iPhone apps, from embedded systems to Windows operating systems. However, being able to write code and develop applications does not mean that you can become an authority on the back end of the APP Store (or improve system efficiency), and being able to develop a real-time multitasking operating system (RTOS) does not mean that you can become an authority on the cost of extending the Windows operating system. Weigh the experts in the field.

Of course, as the core development team of the ProgPow algorithm, IfDefElse is not saying that Windows designers are not "excellent programmers", but it must be explained that because different people have different technical backgrounds, it is easy to cause differences in understanding and understanding of different fields. Assumption bias, especially when discussing the topic of economies of scale.

Similarly, a hardware designer might also dabble in different fields, such as designing a chip for an electric toothbrush, or building a chip architecture (silicon architect) for networking equipment. An engineer who produces 100,000 electric home toothbrush chips may not understand the available economies of scale that a network engineer who produces a million chips thinks about, and a cryptocurrency ASIC chip designer may know very little about GPU-ASIC chip design—these There are not many connections between industries, and some even have gaps between countries.

The last point we want to mention in the overview is that programming and engineering are actually a kind of skill. Unless you are programming and writing code every day, you will soon fall behind and cannot become an authority, because the knowledge update iteration in this area is very difficult. quick. Perhaps this is why it is difficult for new cryptocurrency ASIC manufacturers to enter the mining market based on the SHA-256 algorithm. After all, it is unlikely for a novice programmer to catch up with an engineer who has been working on the SHA-256 algorithm for six years.

Some "hardware bricks" are going out of their way to reassure software engineers that they can beat the cryptocurrency ecosystem -- we've seen this with cryptocurrencies like Monero, Bitcoin, and ZCash already. Appear. But the reality is that this challenge has not happened so far, think about it, if Bitmain or Innosilicon try to make a CPU, do you think they can beat Intel and AMD?

secondary title

Analysis of ASIC chip design cost

Economies of scale are always prevalent—whether it is from a cost perspective or an experience perspective. As for the cost of ASIC chip design, chip designers always seem to have a big debate. Let Mr. Odaily (WeChat: o-daily) take everyone to take a look and analyze the nine issues that have attracted the attention of the industry:

Question 1: Whether the mining algorithm is ProgPow or ETHash, the hash value is determined by the storage bandwidth of the external dynamic random access memory (DRAM), is this true?

But in fact, it's not. The hash value of ProgPow is determined by two factors:

1. Computing core

2. Memory bandwidth

This is why there is a difference between Ethash and ProgPow, as shown in Figure 1 and Figure 2 below:

Figure 1: Comparison of Mining Hash Rates of Nvidia Chip Products

Figure 2: Comparison of Mining Hash Rates of AMD Chip Products

At this stage, ETHash mining is more profitable. The memory demand for this algorithm has increased significantly. The growing demand for high-bandwidth memory has also prompted the development of next-generation high-speed memory technologies, such as GDDR6 (with a bandwidth speed of 768 GB/s ) and HMB2 (bandwidth speeds up to 256 GB/s).

Not all the demand for high-bandwidth memory comes from "Ethash". The entire high-bandwidth memory market is as high as $15 billion, of which only a small part comes from the mining industry. The core market requirements for high-bandwidth memory mainly include: GPU, Field Programmable Gate Array (FPGA), Artificial Intelligence (AI), High Performance Computing (HPC), and games. Compared with the $1.2 trillion artificial intelligence market, the $30 billion PC gaming market, the $35 billion handheld game console market, and the $29 billion high-performance computing market, the high-bandwidth memory requirements of the mining industry are really "negligible".

Question 2: Since ProgPow's existing architecture and algorithm are similar to ETHash, will Innosilicon's next ASIC chip be tailored for ProgPow?

In fact, the only similarity between ProgPow and ETHash is the use of an acyclic graph (DAG) in global memory. From a computational point of view, ETHash only requires a fixed "keccak_f1600" core and a modulo function. ProgPow, on the other hand, needs to be able to execute 16-lane wide random math sequences while also being able to access the high-bandwidth L1 cache. Designing a computational kernel capable of executing ProgPow mathematical sequences is much harder than designing one capable of implementing a fixed-function hash like "keccak".

Another thing to note is that the hash value of ETHash depends only on memory bandwidth, while the ProgPow algorithm depends on both memory bandwidth and the core calculation of random mathematical sequences-this is very important to understand.

The essence of Proof of Work (PoW) is to prove mathematical calculations by consuming hardware and energy costs. As an algorithm, ETHash does not consume most of the hardware costs (computing engines) in mathematical proofs. On the contrary, ETHash only captures the memory interface, which is why you can use an ASIC chip for cryptocurrency mining to eliminate the part of the mathematical calculation that is not captured.

Question 3: Since the GPU is a general-purpose acceleration chip, the cycle of designing, manufacturing and testing the GPU usually takes about twelve months, and a lot of hardware simulation and software development work is required to enable it to cover different computing scenarios and scenarios.

ProgPoW hopes to be able to capture the full hardware cost (as much as possible), because the updated part of the algorithm can capture the computing hardware running different computing scenarios - until the architectural folds (architectural wrinkles) - so it is possible for ASIC chip design It takes more than 3-4 months.

Due to the long time span, another question arises: why is the floating point operation omitted? The answer to this question is actually quite simple: floating-point operations are not portable across chips, and different chips tend to handle edge cases ( corner case). Corner cases, also known as pathological cases, refer to problems or situations in which operating parameters are outside normal ranges, and are mostly cases where several environmental variables or conditions are at extreme values, even if these extreme values ​​are Still within (or bounding) the parameter specification. One of the biggest divergences is the handling of Not-a-Number Values ​​(NaN), which occur naturally when using random inputs, as explained by the Wikipedia page:

If there are multiple Not-a-Number (NaN) inputs, the payload result should come from one of the NaN inputs, but the standard doesn't specify that.

This means that basically every floating point needs to be paired with an "if (is_special(val)) val = 0.0" check if it is going to use floating point arithmetic, which can usually be done in hardware, thus also allowing ASIC chips for cryptocurrency mining benefit from this.

Next, what is the hash rate (Hashrate) and "hash-per-watt"?

Hash rate is a measure of energy cost, and as long as everyone measures it the same way, energy consumption per unit is less important - miners will just keep putting as much energy into mining as possible. But even if you switch the unit of measurement from 1 ETHash (a smaller unit like a joule) to 1 ProgPow-hash (a larger unit like a calorie), the economics of operating costs don't actually change. The Global Hashrate evaluates everyone's total economic weight to secure the network share, and as long as everyone's contribution is measured fairly and in the same units, switching to the ProgPow algorithm will bring no cost to the average miner. What a change.

Of course, some people will say that if Ethereum implements the ProgPow algorithm, it may help to concentrate miners in large mining farms with high-end GPUs, and it will also stimulate mining farms to upgrade GPUs to the latest models. But ifDefElse, the ProgPow algorithm development team, needs to reiterate that economies of scale will always exist, and it is an unavoidable fact in the real world.

Question 4: Compared with GPU, ASIC chip manufacturers can use smaller GDDR6 memory to gain cost advantage. 16 sticks of GDDR6 4GB can achieve twice the bandwidth advantage while maintaining memory cost levels, right?

First of all, having twice the bandwidth advantage requires twice the calculation, which is actually a kind of linear scaling and cannot be regarded as an advantage.

Secondly, we should not be ready to produce 4GB memory chips for GDDR6 yet. The world's third largest memory chip maker Micron (US Micron) only produces 8GB chips, while Samsung produces 8 GB and 16 GB chips. For memory chips, the GDDR6 IO interface area is very expensive, and each generation interface occupies more actual memory dies than memory cells, because the port physical layer (PHY) cannot pass through the process like memory cells. Means shrink.

It is undeniable that what really drives the memory market are some "long-cycle buyers", such as game consoles, GPUs, etc., who also tend to support larger-capacity memory. In fact, today's memory suppliers have no incentive to mass-produce a 4GB memory, after all, the market demand for this memory capacity is not great.

Question 5: There are many modules in the RTX2090 chip that occupy a large amount of chip area and are useless to ProgPow, including PCIE, NVLINK, L2Cache, 3072 slice units, 64 ROPs, 192 time measurement units (TMUs), etc. How do you view this issue?

RTX2080 is not a good reference for discussing this issue. Due to some new features, some modules in Nvidia's RTX series chips occupy most of the chip area, such as ray tracing cores. The ProgPow design is used in conjunction with stock chip products in the Nvidia and AMD ecosystems, so new features in new Nvidia and AMD chip products cannot be used.

If you want a better analogy, maybe AMD RX 5xx series or Nvidia GTX 1xxx series are a good reference. As we mentioned before, there are also some functions in the GPU that are not utilized by ProgPow, such as: floating point logic, level two (L2) cache, and texture cache and ROP, etc. The slice unit is where the vector math is performed, which is definitely what ProgPow requires. ASIC chips for cryptocurrency mining also hope to add regions that enable "keccak" functionality. As the team behind the ProgPow algorithm, we estimate that a ProgPow ASIC chip would have a 30% smaller die area than an equivalent GPU—but would consume only a maximum of 20% less power, even in the best case. In contrast, although some logic modules on the GPU are not fully utilized, resulting in a waste of chip die area, the power consumption is minimal.

Question 6: Will small chips be more profitable than large chips?Compare Logic-Array To ASIC-Chip Cost per Good DieHow to put it, this sounds like popularizing chip manufacturing knowledge, maybe we need to write a "chip manufacturing 101" training document. In addition, for the income calculation formula, you can refer to an article published in 2006 "

", in which you will find that as early as 13 years ago, chip revenue and process control had been greatly innovated.

For a chip with a single functional unit, a chip with a smaller die area will yield higher yields than a chip with a larger die area. But with modern GPUs, this is not the case. Today's GPUs can be recovered and combined almost arbitrarily, and the defects of small replication units are basically negligible. As long as each compressible functional unit is small enough, GPU chips can yield almost as high (or even higher) yields as chips with larger functional blocks.

In order to better explain this concept, we can give a simple brain hole experiment:

1. Suppose you have a large chip "Giant ChipA", which occupies the entire wafer. This "Giant ChipA" is composed of 100,000 detachable sub-assemblies, but it must be ensured that 80% of the sub-assemblies are defect-free to ensure that the "Giant ChipA" works normally, and during the embedding process, bad sub-assemblies will is bypassed.

2. In addition, suppose you also have a small chip "Tiny ChipB", which consists of only one functional module (not embeddable), but this small chip is small enough to assemble 100,000 subcomponents on the same chip. In this case, as long as one subcomponent is broken, it means that the whole "Tiny ChipB" chip is bad.

3. If 20,000 defective subassemblies are evenly distributed on each wafer, then "Giant ChipA" can earn 100% because they can remove 20% of defective subassemblies, while "Tiny ChipB" The return may be only 80% because they cannot remove the defective sub-assembly.

If you look at AMD's Polaris 20-series products and Nvidia's GP 104 products, you'll see these GPUs deployed under a simulation lens consisting of a large number of tiny "detachable" sub-modules.

Question 7: ASIC mining machine voltage can be easily reduced to 0.4V, which is only 1/2 of GPU... Such low voltage ASIC design has been adopted by Bitcoin mining equipment ASIC mining machine manufacturers, so now we have no There is no reason to believe that they will not apply this strategy to ProgPow ASIC miners, can you talk about this issue?

When the chip is only composed of calculations, then the low-voltage design can work, such as an ASIC mining machine dedicated to the calculation of the SHA256d mining algorithm. Integrating other components—such as SRAM, which is also required for ProgPow data caching—is extremely difficult and impossible to operate at low voltages.

Question 8: The same energy-saving effect can also be realized on LPDDR4x DRAM, and its power consumption is lower than that of GDDR6. Let's talk about this issue.

Can't just consider the energy consumption issue, the bandwidth of LPDDR4x is much lower than GDDR6, the bandwidth of each pin of the former is 4.2Gb/s, and the latter is 16Gb/s. The LPDDR4x computing chip requires four times as many memory chips and four times as many memory interfaces to achieve the same performance as GDDR6. In this way, the cost actually increases significantly.

It is worth noting that the interface of high-bandwidth computing chips is usually limited, which means that the chip module area must be large enough, and the surrounding area hardly allows any signal to fall off from the chip to the printed circuit board (PCB). LPDDR4x design requires about Four times the number of chip perimeter pads can achieve the same bandwidth, that is to say, the cost is not only on the memory chip, but also the cost of calculating the chip area, so the total cost is actually not low. . To make matters worse, since any chip is speed-oriented, when the chip module area is larger, it means that the wasted power will also be more.

So, let's think again why today's GPUs can't run on LPDDR4x. First of all, LPDDR4x does not perform as well as expected in terms of bandwidth cost. For a given bandwidth order (four times the number of chips), the cost of LPDDR4x is more than four times higher, which in turn leads to a significant increase in cost - LPDDR4x at 9W The cost of 256 GB/s bandwidth at high power is about 150 US dollars, compared with the same bandwidth cost of GDDR6 at 11W power is less than 40 US dollars, so LPDDR4x does not save miners much money (note that this is about bandwidth cost, not memory capacity cost).

Question 9: GPU manufacturers like Nvidia employ about 8,000 people to develop GPUs, and these GPUs are also very complex; while ASIC manufacturers like LinZhi employ only a dozen people, and only develop GPUs for the ETHash mining algorithm. ASIC miners. The labor costs of these companies differ by 100, so it can be said that ASIC chips have an advantage over GPU chips in terms of cost and time to market.

The point to be made here is that economies of scale are an important factor. The GPU industry is also amortized in various sales channels around the world. The current total market size is about 420 billion US dollars, of which AMD has a market value of about 11.6 billion US dollars, Nvidia is about 154.5 billion US dollars, and the largest Intel is about 254.8 billion US dollars. For the memory market alone, there is also the need to amortize the cost of physical ports (PHYs) and chips in a $500 billion industry, of which Samsung Electronics, with 320,671 employees, has a market value of about $325.9 billion, and they are also in the U.S. The most active patent applicant; the second is Micron Technology with 34,100 employees, its market value is about 60.1 billion US dollars, but the first chip manufacturer to develop 20Gbps high-speed GDDR6 memory; Hynix has 187,903 employees, market value About $56.8 billion, they developed the world's first 1Ynm 16Gb DDR5 DRAM. In comparison, the ASIC chip industry used for cryptocurrency mining has a total market value of only $146 billion, of which $73 billion belongs to Bitcoin.

In addition, we also need to look at the time to market and the Technology Acceptance Model (TAM). Here, we might as well use the development time of the famous S9 mining machine successor as a reference. If it takes three years to iterate the SHA256d algorithm computing chip that has been fully developed and the calculation difficulty is not very high, then what can guarantee that the ASIC mining machine that supports the ProgPow algorithm like the GPU can be put into production quickly? We can also analyze the situation of ASIC mining machines for mining Ethereum cryptocurrency recently. The GDDR6 chip sample trial period has been around for a year, and there is still no new version of the product that can be widely used.

Final Thoughts from IfDefElse, ProgPow Core Dev Team

ProgPow is actually targeting a type of mining hardware that is supported by economies of scale, has high visibility, and has gained a large competitive advantage.