A graphics processing unit (GPU) is a computer chip designed to perform a very large number of simple calculations at the same time. GPUs were built to draw images on screens, where millions of pixels must be shaded in parallel, but the same design suits any job that can be split into many identical small tasks. In money and finance, GPUs matter mainly for two reasons: they were the workhorse of crypto mining for most of the 2010s, and today they power risk models, fraud scoring and artificial intelligence. The term is sometimes written "graphical processing unit", but the industry name is graphics processing unit.
A GPU sits beside the central processing unit (CPU) in a computer. The CPU has a few powerful cores suited to varied, step-by-step work; the GPU has thousands of smaller cores suited to repetition. In mining, GPUs occupy the middle ground between general-purpose CPUs and the single-purpose application-specific integrated circuit (ASIC). Nvidia marketed its GeForce 256, released in 1999, as the first GPU, and Nvidia and AMD remain the two main makers of discrete GPUs.
The GPU in plain terms
Think of a CPU as a small team of senior chefs who can cook any dish, one at a time, very well. A GPU is a hall full of trainee cooks who can each do only simple tasks, such as chopping onions, but who all work at once. If the job is "chop ten thousand onions", the hall finishes long before the senior chefs.
Mining a coin based on proof of work is an onion-chopping job. A miner has to try an enormous number of guesses, each one a quick calculation, until one guess happens to fit the network's target. Because each guess is independent of the others, the work spreads naturally across the GPU's thousands of cores.
How GPUs are used in mining
A proof-of-work miner repeatedly feeds a candidate block header, with a changing nonce, into a hash function and checks whether the result falls below the network target. Each attempt is independent, so a GPU can run thousands of attempts at once. The number of attempts a device makes per second is its hashrate, and the useful comparison between devices is hashrate per watt of electricity, since power is the main running cost.
A typical GPU mining rig links six to twelve graphics cards to one motherboard through riser cables, with an open frame for cooling. Most GPU miners join a mining pool rather than mining alone, because a single rig would wait a very long time to find a block by itself. The pool pays them under schemes such as pay-per-share (PPS) or pay-per-last N shares (PPLNS).
From CPUs to GPUs to ASICs
In the first year of Bitcoin (BTC), people mined with ordinary CPUs using the reference software released by Satoshi Nakamoto. In 2010 developers released GPU mining software, and graphics cards quickly outperformed CPUs by a wide margin, which pushed up mining difficulty for everyone. Field-programmable gate arrays briefly followed, and from 2013 the first Bitcoin ASICs arrived. Because an ASIC is wired to compute only Bitcoin's SHA256 function, it beat GPUs so heavily on hashrate per watt that GPU mining of Bitcoin stopped being profitable.
GPU miners then moved to other coins. Some projects chose ASIC-resistant algorithms that need a lot of memory, on the theory that memory is harder to specialise than raw arithmetic. Scrypt, used by Litecoin (LTC), was an early attempt, although scrypt ASICs appeared from 2014. Ethereum's Ethash algorithm was memory-hard by design and kept GPUs dominant on that network for years.
The Merge and the GPU mining market
The largest single use of mining GPUs ended on 15 September 2022, when Ethereum completed "the Merge" and replaced proof of work with proof of stake (PoS). Ether could no longer be mined, and the hashrate that had secured it moved to smaller proof-of-work coins such as Ethereum Classic (ETC) or left mining altogether. Many used mining cards were resold to gamers and to companies running machine-learning workloads.
The episode showed a risk specific to GPU miners. Their hardware is general-purpose, so it keeps resale value, but the coins it can mine profitably depend on decisions made by each network's developers. A change in consensus rules, a new ASIC or a fall in coin prices can wipe out a GPU miner's margin almost overnight.
GPUs beyond mining
Outside crypto, GPUs are now central to financial computing. Banks and trading firms use them for Monte Carlo simulation in risk and derivatives pricing, and payment companies use them to train the machine-learning models behind fraud detection and transaction monitoring. The same parallel design also makes GPUs useful for password cracking, which is one reason modern password storage relies on a salt and deliberately slow, memory-hard functions.
Demand from artificial-intelligence developers has made high-end GPUs scarce and expensive since 2023, and some former mining companies have repurposed their data centres and power contracts to host AI computing. Critics of GPU mining point to its electricity use and to the electronic waste created when cards wear out or become obsolete; supporters answer that GPUs, unlike ASICs, can be reused for other work.