PrimeGrid: Mapping Prime Numbers with Computers Worldwide

PrimeGrid is not a single hunt for one enormous prime. Through BOINC it runs searches with different mathematical forms and purposes: Cullen, Woodall, generalized Fermat, Proth, factorial, Sierpiński and Riesel problems, and arithmetic progressions of primes. Participants pick a subproject that suits their device, the applications it supports, and the task length they can live with. GIMPS concentrates on Mersenne numbers of the form 2^p−1; PrimeGrid instead keeps a portfolio of prime families and conjecture searches.

At 00:53 UTC on 29 August 2026 the official homepage showed cumulative totals of 357,988 registered users, 893,527 hosts, 101,804 primes found, 38,124 reports to The Prime Pages, and 3,738 mega primes of at least one million digits. Estimated throughput was 3,071.389 TFLOPS. These are not counts of people connected at that moment, and the discoveries do not all carry the same mathematical weight. Cumulative and estimated figures keep moving, so this article timestamps the observation and leaves current values to the official dashboard.

PrimeGrid holds out the possibility that a personal computer will be the first to test an unknown number. A discovery is not a lone winning ticket, though. The people who sieved the candidates, wrote and optimized the applications, checked the computation independently, and ran the servers and databases all stand behind one public record.

The story starts on 12 June 2005. Rytis Slatkevičius opened Message@Home to 50 users from a home laptop while testing a Perl implementation of a BOINC server. Its first experiment, Message7, brute-forced a message encoded with MD5. The name PrimeGrid was chosen on 1 September 2005. Today’s mathematics project did not arrive fully formed; it grew out of a small experiment in getting distributed-computing infrastructure to work.

Early work included the RSA-640 and then the RSA-768 factoring challenges. In March 2006 the project turned away from factorization toward generating and searching for primes. Collaboration with Riesel Sieve and Twin Prime Search, plus the arrival of the LLR application that year, shaped the number-theory project that followed. The old experiment with a cryptographic challenge and the prime searches of today should not be read as one continuous objective.

Challenge Series, PRPNet, generalized Fermat searches, and GPU-enabled Genefer were added over time. A peer-reviewed 2015 paper describes the shift from brute-force factorization of RSA numbers to a large volunteer-computing platform for number theory. BOINC supplies distribution and accounting; on top of it, PrimeGrid builds search plans, specialized applications, validation, and discovery reporting for each family of numbers.

Restricting the form, say to Cullen n×2^n+1, Woodall n×2^n−1, or generalized Fermat b^(2^m)+1 numbers, does more than make unknown primes easier to reach. It opens the way to specialized tests and faster big-integer arithmetic, lets observations be compared with theoretical distributions, and can knock out candidates in open conjectures one at a time. A result can matter because it is the first of its form, because it rules out a k, or because it establishes a new arrangement, even when it is not the largest by digit count.

The first AP27, found on 23 September 2019 after more than three years of searching, shows what that means. Twenty-seven numbers are all prime, and every adjacent pair is separated by the same difference. PrimeGrid is not just a contest for the longest number; it is a set of investigations that survey the terrain of primes in different ways.

Candidate design comes first: pick a mathematical form and an unexplored range, then strip out known factors and earlier results. Sieving cheaply removes candidates divisible by many smaller primes. A deeper sieve leaves fewer survivors, but at some point the sieve costs more than testing the remaining candidates directly. That trade-off is why PrimeGrid’s documentation talks about an optimal sieve depth.

Survivors go through a probable-prime (PRP) test in software such as LLR, PRST, or Genefer. Almost every composite fails here. A pass is extremely promising, but one host showing “prime” is not yet a record. Some long main computations leave data that a short proof task can verify; other work is repeated independently on a second host. Arithmetic-progression searches and sieves do not all take the fast-proof path.

Where required, another program or a method such as N−1 then supplies a mathematical primality proof, and The Prime Pages runs its own checks when it registers a result. So “every task is always computed twice” and “one proof task proves everything mathematically” are both inaccurate. Error detection, reproducibility of the computation, and proof of primality are combined differently for different kinds of work.

Home machines face overclocking, memory errors, drivers, unstable power, and software faults. PrimeGrid spends resources on computational proofs and double checks not simply because volunteers are distrusted, but because honest hardware can slip during a long big-integer calculation. A speed ranking and reproducibility are separate goals.

“Probable prime” does not mean fake. The right tests are designed so that a composite passing them is extraordinarily unlikely, and PRP tests are indispensable to discovery. World records and mathematical claims still record the probable or proven status, the program, the hardware, the checker, and the date, so readers can tell the levels of evidence apart.

PrimeGrid’s official application list includes implementations for x86 and Arm CPUs, NVIDIA CUDA, AMD and Intel OpenCL, and Apple platforms. Not every subproject can use every device. The candidate form, the size of the integer, the FFT or multiplication method, memory, double precision, drivers, checkpointing, and validation together decide which hardware fits. There is no table in which owning a GPU makes every task faster than a CPU.

LLR and LLR2, PRST, Genefer, OpenPFGW, AP-specific code, and sievers such as sr2sieve and AthGFNSieve play different roles. A sieve throws out many candidates with light operations; a PRP test puts one survivor through long big-integer arithmetic; a proof needs different data and different algorithms. Even on one host, changing the application or the size of the number moves the bottleneck among arithmetic units, memory, communication, and heat.

A February 2025 generalized Fermat prime, 13520762^524288+1, was found as a PRP with Genefer on a GeForce RTX 3060 Ti, then confirmed with LLR on a Ryzen 9 7950X3D in about 20 hours 40 minutes. It is a concrete case of GPU search and CPU confirmation using different methods inside a single discovery. The hardware names are only one part of a chain that also runs through software authors, sievers, server operators, and checkers.

On 12 October 2025 PrimeGrid found 2524190^2097152+1. It is the first GFN-21 prime and has 13,426,224 digits. In August 2026 The Prime Pages recorded it as proven and sixth among all known primes. Its N−1 verification there took 78.12 days. The gap between finishing the discovery computation and recording a proven result is where the long verification stage becomes visible.

The 2016 result 10223×2^31172165+1 carries a different kind of value. It has 9,383,761 digits, but the point is that it eliminated k=10223 from Seventeen or Bust. It did not just add a giant number; it removed one candidate from the finite open set around the smallest Sierpiński number.

The 2019 AP27 found 27 primes that begin near an 18-digit value and are separated by an equal interval. Each number is far smaller than a million-digit record, yet the 27-term arithmetic progression was a world first. Digit count, special form, elimination of a conjecture candidate, and a first arrangement do not belong in one league table.

BOINC credit is an accounting of the computation a participant returns. PrimeGrid ranks participants, teams, countries, computers, and prime finders. Alongside credit badges for individual subprojects sit discovery badges for mega primes, for eliminating conjecture candidates, for AP26, AP27 and AP28, and for world firsts. Challenge Series and Tour de Primes add timed contests for individuals and teams, which gives the search a season and a story.

None of that conflicts with contributing without financial compensation. Credit is not fiat money and not a cryptoasset, and it confers no ownership of a research result. Rankings can sustain motivation and a sense of belonging; correctness is settled separately, by validators and proofs. A participant with high credit cannot vote a number into being prime.

The prime-reporting policy lays out how the first finder is contacted and, if no reply arrives, how the double checker is contacted and the result is eventually registered anonymously. Behind a finder’s name stand the people who prepared the range, ran the sieve, and checked the result. Keeping public results and program names on the record turns a leaderboard into a research history others can inspect.

Primes can look irregular and still obey known structure: the density described by the prime number theorem, restrictions from congruences, and the positions they can occupy inside special formulas. The Cullen, Proth, generalized Fermat, Sierpiński, and Riesel numbers that PrimeGrid studies all have a form. Knowing the form makes specialized algorithms possible; it does not yield a formula whose every value is prime, and it does not reveal the secret factors of arbitrary cryptographic keys.

The 2002 AKS result showed that deciding whether an integer N is prime can be done deterministically in polynomial time. That was a major theoretical breakthrough, but it is not a way to recover the secret p and q efficiently from an RSA public value N=p×q. “Is this number prime?” and “Which primes made this product?” take different inputs and ask for different answers.

So “a pattern in primes would bring down the cryptographic fortress” is not a factual statement without a condition attached. What would change cryptography is a practical method for factoring arbitrary deployed keys, a practical discrete-log algorithm, a way to predict secret values from biased key generation, or an attack that exploits the structure of particular parameters. An elegant distribution law, or one more enormous prime, does not supply any of that.

Large-prime searches advance number theory, big-integer arithmetic, error detection, distributed verification, and hardware optimization. Their contribution to cryptography is indirect, in that sense. Discovering a public record prime with a special form does not change the security level of deployed RSA keys or Bitcoin keys. Key strength depends on randomness, parameters, protocol, implementation, and operations, not on bit length alone.

Prime research and cryptography are not unrelated, either. An algorithm that makes factorization or discrete logarithms practical, bias in the primes chosen by a weak key generator, or precomputation against reused parameters can all change security. The question that decides it is not whether both fields use the word “prime,” but whether a discovery hands an attacker a practical route to the secret value.

FIPS 186-5 requires the p and q of an RSA signing key to be generated as random probable or provable primes meeting set sizes and conditions. Public million-digit record primes are not used as secret factors. An RSA-2048 modulus is about 617 decimal digits and each factor is roughly half that length, but comparing digit counts does not settle security strength. PrimeGrid’s huge primes are public research results, not secret-key material.

Real failures have happened when weak randomness left different RSA keys sharing a prime factor. Heninger and colleagues compared large populations of public keys and found cases where taking the greatest common divisor of two moduli recovered the shared factor, and with it the private keys. That was not a mysterious pattern running through all primes; it was implementation and entropy failing to pick independent, unpredictable p and q.

In finite-field Diffie–Hellman, short and widely reused groups can also invite precomputation. The lesson from the Logjam research and RFC 7919 is about group size, structure, provenance, reuse, and implementation, not about the prime being public. Separating public parameters from secrets and naming the actual hard problem is more accurate than the slogan “a larger prime is safe.”

Bitcoin’s ECDSA and BIP 340 Schnorr signatures use the secp256k1 elliptic curve. The field prime p, the curve equation, and the generator G are public parameters. The secret is x and the public key is Q=xG. Security rests mainly on the difficulty of the elliptic-curve discrete logarithm, not on factoring an RSA modulus. A new large prime found by PrimeGrid does not automatically strengthen the secp256k1 parameters or anyone’s existing keys.

Bitcoin mining is not a prime search either. Miners vary the nonce and other parts of a block header, trying inputs until the double-SHA-256 hash falls below the network target. A PrimeGrid PRP test asks whether N is prime; mining asks whether the hash of an input meets a threshold. The huge space of independent trials may look similar, but the mathematical problem, what a result means, how it is validated, and what it pays are all different.

Shor’s algorithm, on a sufficiently large and fault-tolerant quantum computer, gives polynomial-time routes to both integer factorization and discrete logarithms, which threatens RSA, DH, and ECC. That is an algorithmic breakthrough, not “finding a pattern in primes.” No published device has recovered Bitcoin private keys in practical time, so the research should not be turned into an invented deadline. Migration does take years, though, as NIST’s post-quantum standardization shows, which is why early design work remains valuable.

Taking part starts with installing the official BOINC client, attaching PrimeGrid, and choosing subprojects and resource limits. Check the official application list for operating-system, CPU or GPU, and driver support. Set CPU share, schedules, battery, temperature, fans, and networking conservatively at first. Some tasks run for days or longer, so look at deadlines and checkpoints, and never use a device without its owner’s permission.

Spare computation still draws power and gives off heat. The marginal impact depends on whether the hardware was bought for the work, whether its heat is useful, and what the local electricity mix looks like. Overclocking that introduces errors can look fast while adding recomputation. Treat valid results, electricity, hardware life, and noise as one system to run, rather than optimizing for credit alone.

Nothing guarantees that you will find a prime. A composite result still pushes the searched range forward, sieving saves later work, and a proof checks someone else’s result. The appeal of PrimeGrid does not belong only to the lucky host. It lies in ordinary computers around the world filling in a verifiable map one failed candidate at a time.