RandomX is a proof-of-work (PoW) algorithm that is optimized for general-purpose CPUs. RandomX uses random code execution (hence the name) together with several memory-hard techniques to minimize the efficiency advantage of specialized hardware.
RandomX utilizes a virtual machine that executes programs in a special instruction set that consists of integer math, floating point math and branches. These programs can be translated into the CPU's native machine code on the fly (example: [program.asm](doc/program.asm)). At the end, the outputs of the executed programs are consolidated into a 256-bit result using a cryptographic hashing function ([Blake2b](https://blake2.net/)).
Both modes are interchangeable as they give the same results. The fast mode is suitable for "mining", while the light mode is expected to be used only for proof verification.
The first audit was generously funded by [Arweave](https://www.arweave.org/), one of the early adopters of RandomX. The remaining three audits were funded by donations from the [Monero community](https://ccs.getmonero.org/proposals/RandomX-audit.html). All four audits were coordinated by [OSTIF](https://ostif.org/).
Final reports from all four audits are available in the [audits](audits/) directory. None of the audits found any critical vulnerabilities, but several changes in the algorithm and the code were made as a direct result of the audits. More details can be found in the [final report by OSTIF](https://ostif.org/four-audits-of-randomx-for-monero-and-arweave-have-been-completed-results/).
RandomX is written in C++11 and builds a static library with a C API provided by header file [randomx.h](src/randomx.h). Minimal API usage example is provided in [api-example1.c](src/tests/api-example1.c). The reference code includes a `randomx-benchmark` and `randomx-tests` executables for testing.
* The key `K` is selected to be the hash of a block in the blockchain - this block is called the 'key block'. For optimal mining and verification performance, the key should change every 2048 blocks (~2.8 days) and there should be a delay of 64 blocks (~2 hours) between the key block and the change of the key `K`. This can be achieved by changing the key when `blockHeight % 2048 == 64` and selecting key block such that `keyBlockHeight % 2048 == 0`.
The table below lists the performance of selected CPUs using the optimal number of threads (T) and large pages (if possible), in hashes per second (H/s). "CNv4" refers to the CryptoNight variant 4 (CN/R) hashrate measured using [XMRig](https://github.com/xmrig/xmrig) v2.14.1. "Fast mode" and "Light mode" are the two modes of RandomX.
Note that RandomX currently includes a JIT compiler for x86-64 and ARM64. Other architectures have to use the portable interpreter, which is much slower.
Efficient mining requires more than 2 GiB of memory, which is difficult to hide in an infected computer and disqualifies many low-end machines such as IoT devices. Web mining is infeasible due to the large memory requirement and the lack of directed rounding support for floating point operations in both Javascript and WebAssembly.
RandomX uses only operations that are guaranteed to give correctly rounded results by the [IEEE 754](https://en.wikipedia.org/wiki/IEEE_754) standard: addition, subtraction, multiplication, division and square root. Special care is taken to avoid corner cases such as NaN values or denormals.
RandomX generates multiple unique programs for every hash, so FPGAs cannot dynamically reconfigure their circuitry because typical FPGA takes tens of seconds to load a bitstream. It is also not possible to generate bitstreams for RandomX programs in advance due to the sheer number of combinations (there are 2<sup>512</sup> unique programs).
Sufficiently large FPGAs can mine RandomX in a [soft microprocessor](https://en.wikipedia.org/wiki/Soft_microprocessor) configuration by emulating a CPU. Under these circumstances, an FPGA will be much less efficient than a CPU or a specialized chip (ASIC).
