41-million-digit Prime Crunched By Datacenter GPUs

A former Nvidia engineer has found the largest known prime number – a whopping 41 million digits long – using an A100 GPU made by his previous workplace to do the grunt work.

This wasn't the CPU-heavy hunt for primes that we've typically seen. No, this time GPUs, ones more widely used for the datacenter, did all of the heavy lifting, leveraging their high levels of parallel processing power to crack the calculations that CPUs just can't keep up with.

The latest addition, named M136279841, belongs to a special class of prime numbers that are especially large. Mersenne primes equal 2ⁿ - 1, where n is the exponent needed to generate the prime, used to form the name. M136279841, for example, equals 2^136,279,841 - 1. (It's a bit snappier using the M moniker rather than the full prime number – but for those who fancy, the folks at the Great Internet Mersenne Prime Search have a ZIP archive containing all 41,024,320 digits.)

Part of the GIMPS search, which has been going since 1996, and which provides free software for would-be math sleuths, this discovery shows just how much muscle GPUs are bringing to high-performance computing and are useful far beyond just pushing pixels in games or running CAD simulations.

The man behind the 41-million-digit prime number is Luke Durant, a 36-year-old researcher and ex Nvidia staffer. Back in 2023, he began contributing to GIMPS, and after around a year, boom, an Nvidia A100 graphics card located in Dublin, Ireland reported the massive number was "likely" prime. Durant is said to have recognized the potential of GPUs in the search for Mersenne primes and had developed an infrastructure to run the free GIMPS software on many GPU servers, the org said.

To confirm the calculations, an Nvidia H100 located in San Antonio, Texas primed out using a Lucas-Lehmer test. It's also the first time that GIMPS prime has been discovered using a probable prime test, which posed an issue about the official discovery date amongst math boffins, who've already been talking about the discovery for weeks.

Finding M136279841 may not seem like it'll change your daily life, but here's the thing: it's solid proof of just how much compute power we've been sitting on with modern GPUs.

Durant spotted the 52nd Mersenne prime over five years after his predecessor, 35-year-old Florida man Patrick Laroche, found number 51 while running the free GIMPS software on a four-core Intel Core i5-4590T processor over 12 days.

• A quick guide to tool-calling in large language models
• China’s quantum* crypto tech may be unhackable, but it's hardly a secret
• Babbage boffin Ada Lovelace honored for computer science contributions
• Vietnam finds a formula – C = SET + 1 – to grow its semiconductor industry to $100bn

Cryptography, AI, high-performance computing and industries are already capitalizing on using the compute power of graphics over CPUs. While the number itself may not have immediate real-world applications, the fact that it was done using datacenter level graphics cards shows that the future of running compute on just CPUs, even with insane core counts, dual sockets, and more, isn't the only way to get the job done. While AI is the buzzword that just keeps on buzzing, graphics compute isn't just useful for high-performance inferencing, but can also out-compute the actual compute (CPU).

So what's the upshot? We're seeing plenty of breakthroughs and innovations thanks to distributed computing and the raw power of graphics cores. Given that we're expected to see a large uptick in graphics compute capabilities with the next generation cards from Nvidia on the horizon, it's another example of what graphics compute is capable of. Expect more graphics cards in the realms of cryptography and high-level maths - just don't expect to be cracking 41-million-digit prime discoveries with your run-of-the-mill Intel Core i9-14900K and Nvidia RTX 4080 gaming rig any time soon. ®