RandomX algorithm: the Monero network, hardware and specs

RandomX is the proof of work of the Monero network, switched on by a hard fork on 30 November 2019 in place of the last version of CryptoNight. The author of the algorithm is tevador, a notable contribution to the design came from SChernykh, and the original idea of executing random code belongs to hyc. The code is open under the BSD license.

It was conceived with one specific goal: to make a dedicated chip pointless. Before 2019 Monero fought hardware crudely, changing the algorithm roughly every six months, and that regime wore everyone out, exchanges and wallet developers included. RandomX was meant to close the question for a long time, and, jumping ahead, it did close it: a real chip for this algorithm has still not been built, and the machines declared as dedicated ones never got as far as volume shipments.

Hence the main peculiarity of this page. Under RandomX the catalog holds exactly one machine, the Antminer X5, and it is sold as an ASIC for Monero. In reality it has ordinary general purpose server processors inside. Monero developers took it apart in the autumn of 2023 and stated the conclusion plainly: this is one hundred percent not an ASIC. Below we go through what that means for a buyer, and why the one calculation in our series that does not go deep into the red is still no reason to celebrate.

An ordinary proof of work algorithm drives data through a fixed sequence of operations. The simpler and the more predictable that sequence is, the more worthwhile it becomes to cast it in silicon. RandomX breaks that logic at the very foundation: the sequence of operations here is not fixed, it is generated afresh for every hash.

A random program for a special virtual machine is born from the seed. Its instruction set is deliberately close to that of an ordinary processor: integer arithmetic, floating point computation to the IEEE 754 standard with addition, multiplication, division and square root, and branching as well. The program is not interpreted, it is translated into machine code right at run time and launched on a real processor. The result is folded by the Blake2b function into a 256 bit hash.

Then comes the part everything was started for. Mining needs a dataset of 2080 mebibytes, and it has to sit in memory in full. There is a light mode at 256 mebibytes, it gives the same result but computes far more slowly, and it suits verifying blocks from others rather than finding your own. On top of that the algorithm needs hardware AES and two mebibytes of third level cache for every thread.

Put that together and you get a portrait of an ordinary server processor. On a graphics card the algorithm computes badly: the requirement of a little over two gigabytes of memory cuts off cheap devices, and dense branching sits poorly on a parallel architecture. On programmable logic it does not compute at all: loading a new configuration takes tens of seconds, and the number of program combinations here is astronomical, so compiling them in advance is impossible. The seed meanwhile changes every 2048 blocks, roughly once every 2.8 days.

1. Epoch seedChanges once every 2048 blocks, sets the set of programs
2. A random programGenerated for every hash, translated into machine code
3. A 2080 mebibyte datasetSits in memory in full, accessed at every step
4. ExecutionIntegers, floating point, branching, hardware AES
5. Blake2b foldingA 256 bit result, compared with the target

• 2019Before this year Monero changes the algorithm roughly every six months to devalue the dedicated hardware that keeps appearing. The regime wears out the whole ecosystem.
• 2019In March the developers work through eleven variants of what to do next, from speeding up the algorithm changes to moving to a deliberately ASIC friendly hash. They pick RandomX.
• 2019On 30 November a hard fork switches RandomX on. The algorithm change every six months stops.
• 2022In May tail emission begins: the block reward no longer falls and stays at the level of 0.6 XMR. The reward will never drop to zero.
• 2023On 28 August Bitmain announces the Antminer X5 and calls it the first ASIC for Monero. Sales start in September.
• 2023In September Monero developers take the X5 apart and publish their conclusion: it is not an ASIC, there are general purpose server processors inside. They did not go through with an algorithm change.
• 2025In the summer the Qubic pool builds up its share of hashrate and claims control over the majority. A look back afterwards shows a different picture, details below.
• 2025In December Bitmain announces the Antminer X9 for RandomX: 1 MH/s at 2472 watts, that is 2.47 J/kH. In May 2026, according to reports from buyers, the model is pulled from sale before the first shipments, with prepayments returned.
• 2026On 16 January the network sets its all time high of hashrate, 7.54 GH/s. By the summer of 2026 the network drops back to roughly 5.5.

The only catalog model on RandomX. Hashrate in kilohashes per second, efficiency in joules per kilohash. Click the name to open the profitability calculator.

There is no custom firmware for the Antminer X5. AsicBoost firmware is built for SHA-256 boards, on Scrypt two models are supported separately, and there are no RandomX machines on that list. Network installation through HashCore Toolkit belongs to those same supported lines.

Third party projects show the same picture. We went through the official supported model lists of the known firmware developers: the X5 is in none of them, and no machine on RandomX is listed there at all. All those projects work with SHA-256 and partly with Scrypt.

The reason is technical, and it is a curious one. The custom firmware ecosystem is built on overclocking and undervolting hash boards with dedicated chips. The X5 has no hash boards at all: it holds server processors running Linux and ordinary memory modules. There is nothing to overclock in the usual sense.

The developers who took the machine apart noted that with the factory limits removed it would come out as a general purpose compute node. No public project doing that exists, and buying the machine in the hope of such an option is not worth it.

This is the only page in our series where the calculation does not go deep into the red, and that is exactly why it has to be read carefully. At the Spanish tariff of about 0.15 euro per kilowatt hour the Antminer X5 burns roughly 146 euro of electricity a month and mines roughly 151 euro worth. The result is about five euro in the black over a month. That is not profitability, that is zero within the margin of error.

The break even threshold on the tariff works out at about 0.155 euro per kilowatt hour, that is exactly at the level of the ordinary Spanish price. There is no headroom at all. The XMR rate has to hold above roughly 337 dollars, otherwise the machine goes into the red. If network hashrate returns to the January 2026 high, and that was almost one and a half times above the present level, the same X5 at the same tariff will be losing about 35 euro a month. There is nothing to say about the machine itself paying off at such figures: at five euro in the black a month it will never pay off.

And one more comparison, an awkward one for the seller. A dual socket server on modern server chips gives roughly ten percent less hashrate at a comparable power draw. In money terms it comes out no worse than the X5 under the same conditions, and sometimes better. The difference is that after an algorithm change the server stays useful hardware, while the X5 stays a box for one coin.

Open the profitability calculator →

Efficiency here is measured in joules per kilohash, J/kH, and this is the only page in our series with such a unit. The reason lies in how the algorithm is built: every RandomX hash is the execution of a whole random program with accesses to a little over two gigabytes of memory. One such hash costs roughly a billion times more than one SHA-256 hash, which is why the hashrate figures here are nine orders of magnitude smaller. Comparing 212 kH/s with the hundreds of TH/s of Bitcoin machines is meaningless.

There is one model in the catalog, so there is nothing to compare within the page. What is useful instead is a comparison with ordinary hardware. A top home processor gives roughly seven and a half times less hashrate than the X5 and is about a third worse in efficiency. A dual socket server on modern server chips falls behind the X5 by only a little over ten percent in hashrate and is roughly equal to it in efficiency. That is the measure of how far RandomX succeeded in its task: a real ASIC usually beats a processor by hundreds and thousands of times, and here the gain is measured in tens of percent.

The arithmetic goes as usual: multiply efficiency by hashrate in kilohashes to get draw in watts, divide by 1000 to get kilowatts, multiply by your tariff and by 720 hours a month. For the X5 that is about 146 euro a month at the Spanish tariff. Compare the sum you get with the mining forecast in the calculator, and here it is especially important to take a fresh rate and a fresh network hashrate: on this machine the difference between the black and the red fits into a few percent.

Count at your own tariff, not at the calculator default. Most calculators show profitability at an electricity price half the Spanish one. At 0.15 euro per kilowatt hour the X5 balances right on the edge: the break even threshold is about 0.155 euro. An error of one cent changes the sign of the result.

Build in the risk of an algorithm change. For Monero that is not a hypothesis but the working practice of past years. An update to RandomX itself was released in the spring of 2026 and is waiting for activation in the network, and back in 2023 the developers estimated that a small edit would knock down the performance of the processors sitting in the X5 noticeably harder than the performance of ordinary chips. At a margin of five euro even a drop of twenty percent makes the machine unprofitable.

Compare it with an ordinary server. A dual socket server puts out comparable hashrate at comparable efficiency, costs roughly the same money and stays useful hardware after any algorithm change. For RandomX that is a fair alternative, not a stretch.

Think through where you will sell what you mine. XMR liquidity is fragmented after delistings from a number of large venues, and a noticeable share of turnover falls on one pair on one exchange. From July 2027 European legislation will restrict the work of regulated venues with coins of enhanced anonymity. That concerns venues, not ownership, but the usual route to selling in the EU will be narrowing.

Check the conditions where you will put it. 1350 watts and 75 decibels. For a home in Spain that means a separate line, a separate room and a conversation with the neighbours. The machine is not quiet.

Do not prepay for batches that have not shipped. Announcements of machines several times better in efficiency turn up around the algorithm from time to time. Official confirmation is no insurance here: under RandomX it has already happened that a machine announced with a price and dates was pulled from sale before the first shipments, with prepayments returned.

Taking the Antminer X5 for a real ASIC. Inside sit general purpose server processors and ordinary memory. Monero developers took the machine apart and stated this plainly. In efficiency it is comparable to a dual socket server, so the hundredfold gap usual for an ASIC is not there.

Comparing kilohashes with terahashes. The figure of 212 kH/s next to the hundreds of TH/s of Bitcoin machines looks laughable, but the units are not comparable: on RandomX every hash costs roughly a billion times more. What has to be compared is income, not bare numbers. The reverse mistake also turns up: listings confuse kilohashes with megahashes, a difference of a thousand times.

Ignoring the risk of an algorithm change. Before 2019 Monero changed the algorithm roughly every six months precisely in order to fight hardware. An update to RandomX was released in the spring of 2026 and is waiting for activation in the network. A SHA-256 machine has no such risk, this one does.

Using a tariff that is not yours. At the electricity price that calculators substitute by default the machine looks profitable. At the Spanish 0.15 euro per kilowatt hour it balances around zero. Different calculators at one and the same tariff give results of opposite sign, because they take a different rate and a different network hashrate.

Counting on a fallback coin. Wownero uses a different algorithm and is incompatible. Zephyr and Salvium formally fit, but these are networks orders of magnitude smaller: everything Salvium mines in a day is measured in hundreds of dollars. In money terms both are worse than Monero.

Waiting for a custom firmware with overclocking. No known project has one, and the technical basis for it is missing: the machine has no hash boards of the kind that usually get overclocked and undervolted.