The bitcoin mining industry has witnessed massive change over the past two years. The technological arms race launched by ASIC makers quickly put an end to GPU and FPGA (field-programmable gate array) mining, but much like the Cold War arms race, additional investments may prove unsustainable in the long run due to ROI constraints.

Currently, miners are hitting the wall. Technology is the first problem. Most miners are already using the latest nodes, namely 28nm and 20nm processors offered by TSMC and GlobalFoundries, while the first 14nm/16nm FinFET ASICs are expected next year. Yet progress is slowing down due to a number technical limitations plaguing all chipmakers.

The second problem involves economics. It is more down to earth, but it is closely related to chip design and manufacturing. Bigger chips manufactured on relatively immature processes tend to be costlier to produce and develop. They usually face yield and leakage issues as well.

The first technical challenge can be described as the thermal barrier. Simply adding more transistors and building massive chips with billions of transistors delivers more performance, but cooling these chips becomes troublesome and impractical. At the same time efficiency becomes an even bigger problem.

As an emerging field of IC design, bitcoin mining ASICs have experienced rapid evolution over the past two years. However, they cannot keep evolving and developing at the current rate. A number of technical limitations and Moore’s Law simply cannot be overcome even by the world’s foremost chip designers and foundries.

The same issues apply to multibillion dollar chipmakers like Intel, Qualcomm, Nvidia, AMD and various ARM partners such as Samsung, MediaTek and Apple.

Moore’s Law stipulates that the number of transistors doubles approximately every two years. Performance scales accordingly, but this does not mean that performance increases at a linear rate, as chip designers can optimise microarchitectures without piling on more transistors, hence yielding more performance from a the same number of transistors and chip area.

ASICs differ from general purpose processors and a lot less money goes into its development and optimisation. As it matures, optimisation gets more difficult. This means ASIC design will inevitably slow down.

Since most ASICs already use cutting edge manufacturing nodes and there is a limit to how big practical designs can be, more resources will have to go into optimisation, performance-per-Watt tweaks and state of the art cooling technology. This approach involves more spending and development than a transition to a new manufacturing process and it usually does not yield the same performance or efficiency increase.

ASIC designers tend to keep a lot of information away from prying eyes. ASIC makers reveal some basic specs, such as the number of processing cores and the size of the chip package, but they do not paint the full picture. The actual number of transistors and the size of the actual silicon in the package (die size) tend to remain undisclosed, although they are much more relevant than package sizes, voltages and other marketing-friendly specs. Since these are not mass market chips, they are unlikely to be taken apart and placed under a microscope so much of the information never becomes public, thus preventing in-depth analysis.

The problem for bitcoin miners is the fact that overinvestment is starting to create an unsustainable trend. Demand, caused by ever higher difficulty, is outstripping development. Ideally miners would need chips capable of breaking Moore’s Law and then some. Over the past year the hash rate of the bitcoin network has skyrocketed from around 1,000,000GH/s to more than 200,000,000GH/s, briefly peaking at 231,138,370GH/s in late August. In roughly the same period the difficulty shot up from about 65 million to 27,428,630,902 on 31st August.

Part of the increase should be attributed to new, more efficient and powerful ASICs, but in reality most of it comes from huge investments in this inherently risky industry. The old approach no longer works, as illustrated by hash rate trends in late August and early September. This is the focal point; this is where technology and economics intersect.

Since Moore's Law is not going anywhere, the current rate of growth is unsustainable unless the price starts to pick up. It can be maintained through additional investment in industrial-scale mining operations, but only in theory. The days of high yields and ROI measured in weeks rather than months are over. Technological advances, including the introduction of more efficient ASICs and the transition to 14nm/16nm FinFET nodes have the potential to fuel further growth, but not nearly at the same rate.

Technology simply cannot evolve at a rate that would enable such growth in the long run without additional investments.

Thermals and efficiency are becoming a big problem. At a price of $500 bitcoin miners generate about $1.8m in revenue each day. If the price does not go up they could end up making even less. If it goes down, pulling the plug is another option, as miners will run their hardware at a loss for long. This figure does not include operating expenses, cost of capital and investments in next generation hardware. Energy costs are another constant. They are more likely to go up than down, forcing miners to migrate to regions with abundant, cheap electricity. Iceland and Scandinavia have already attracted a number of mining operations. This trend will inevitably lead to even more centralisation.

The cost of keeping the network running is going up, but the returns are not. Margins are going down, competitiveness is becoming increasingly important and small mining operations are unlikely to remain competitive in the future. This does not include just individual miners, but small mining companies as well. Therefore we expect to see more consolidation and diversification moving forward.

So will the network continue to grow more powerful? There is no straightforward answer as nobody can estimate the price of bitcoin over the next few quarters.

Should the price remain stagnant, we could even start to see a drop or at least a stagnation in the hash rate. The difficulty has gone up three times last month, going up from 19,729,645,941 to 27,428,630,902 in under 25 days. This represents an increase of more than 40%. Each increase was followed by a sharp (but brief) decline in the hash rate, which was compensated for in the days following each increase. However, the past two weeks point to a different trend.

The increase also indicates that many miners responded to the difficulty spike by pulling the plug on obsolete hardware, which itself is nothing unusual. However, the size of each drop is, as we are dealing with drops in the range of 20-25%. Big difficulty increases past the 20,000,000,000 mark clearly affected a large part of the network, rendering many miners obsolete overnight.

They were replaced by more efficient units that kept the hash rate up, but in turn they also increased the difficulty, creating a vicious circle that will claim even more obsolete hardware in the near future. This is nothing new, but data indicates the sheer pace at which the difficulty is going up is starting to have a serious effect on many miners. In the short term it will create market opportunities for the most competitive mining operations and hardware makers, boosting efficiency in the process. The risk is that the trend will claim too much obsolete hardware and put many miners out of business, resulting in even more centralisation and fewer incentives to invest in the mining space.

Michael Bedford Taylor, Bitcoin and the Age of Bespoke Silicon, 2013

Mining hardware stats are notoriously hard to come by, so we used the most visible, crowdsourced dataset available, which is Bitcoin.it's mining hardware comparison data. We included miners that are listed as discontinued or currently shipping from the Bitcoin.it data. When there was no listed shipping date for a miner on the bitcoin.it table, we searched for shipping dates from publicly available sources, particularly miner's own websites or their posts on Bitcoin Talk. We excluded the three Minerscube miners because there is debate about the units' performance, as pointed out in the comments below.

Data and charts by Joon Ian Wong.

Image via Shutterstock