MSTR Cat

Blockchain/Level III · The network/3.2 Mining/Lesson 1

From CPUs to GPUs, FPGAs and ASICs

The hardware arms race: from general-purpose logic to single-function silicon

MaterialsManufacturer specifications for an Antminer S9 and a current-generation ASICA local electricity rate (cost per kWh)

Wonder

  1. Why does a graphics card, designed for video games, outperform a modern CPU at hashing?
  2. What happens to the economic viability of a mining rig when its efficiency lags behind the next generation by just 20%?
  3. How does concentrating manufacturing in a few firms change the decentralization profile of the network?

Explore

  1. In early 2009, mining was performed on standard Central Processing Units. Satoshi and the first adopters ran software on personal desktops, brute-forcing SHA-256 hashes to compete for block rewards. The era lasted into 2010; CPUs are built for varied, sequential work, not the endless parallel repetition that proof of work rewards. As network difficulty rose, the computational gap between hobbyist hardware and the expanding network widened rapidly.
  2. During 2010, the center of gravity shifted to Graphics Processing Units. Originally built for rendering graphics, GPUs contain thousands of cores capable of performing simple calculations simultaneously. Laszlo Hanyecz was among the first to mine on a graphics card, recognizing that this parallelism offered a massive boost in hashrate per watt compared to CPUs. Within months CPU mining was uncompetitive as difficulty climbed.
  3. Next came Field-Programmable Gate Arrays around 2011. Unlike fixed GPUs, FPGAs are silicon that can be reprogrammed by the user after manufacturing. Miners used them to strip away all unnecessary logic from a computer chip, leaving only the specific circuits needed for SHA-256 hashing. This resulted in significantly better energy efficiency and performance than general-purpose graphics cards, marking another step toward specialized hardware.
  4. The pinnacle of this evolution is the Application-Specific Integrated Circuit. The first bitcoin ASICs shipped in early 2013 (Avalon), and Bitmain’s Antminer line followed; the Antminer S9 (2016), at about 12-14 TH/s depending on batch, dominated for years. An ASIC is a custom-built chip optimized exclusively for the double SHA-256 algorithm, offering unparalleled performance and efficiency that renders all previous hardware types economically unviable in competitive environments.
  5. Efficiency is the key measure: joules per terahash. The best new machines in 2025-2026 run at roughly 10-15 J/TH, while the fleet average is closer to 20-30 J/TH. Each generation of hardware made the previous one unprofitable wherever electricity costs were not negligible. A bitcoin ASIC can do nothing but compute double SHA-256, so it can mine only coins that use the same proof of work. The specialization ties miners’ capital to a single asset’s health.
  6. ASICs concentrate manufacturing in a few firms, a recurring decentralization concern. While the software remains open and permissionless, the hardware required to participate economically is proprietary and produced by limited suppliers. To check current efficiency standards, visit a block explorer or mining dashboard to compare the estimated network hashrate with the power consumption metrics of leading ASIC models from different vintages.

Key facts

  • CPU mining dominated 2009 but was outpaced by GPU parallelism in 2010.
  • Laszlo Hanyecz was among the first to mine using a graphics card in 2010.
  • FPGAs emerged in 2011, offering superior efficiency over GPUs via reprogrammable logic.
  • The Antminer S9 (2016) delivered about 12-14 TH/s and dominated for years.
  • The best recent ASICs use roughly 10-15 joules per terahash; the fleet average is higher.
  • ASIC manufacturing is concentrated in a few firms, raising decentralization concerns.

Exercise nine, six, three, one

9CollectList 9 computing hardware types used or proposed for Bitcoin mining from 2009 to present. Note their primary design purpose and relative efficiency era.
6ClusterGroup the 9 into 6 categories based on architectural specialization: general-purpose, parallel graphics, reprogrammable logic, fixed-function ASICs, experimental, or obsolete.
3DistillReduce the 6 to 3 reasons why each successive generation rendered the previous one economically unviable.
1CrystallizeWrite 1 sentence on what specific technical advantage allowed each new hardware class to displace the last.

Fact check

Dates, figures and claims checked against the published sources listed below.

Open in the interactive academy →

PT ES ZH JA KO RU DE FR MS