
FP64
FP64 is a numeric format for processors and graphics chips that stores each number using 64 bits, providing very high computational precision. It plays a minor role in AI development, but is indispensable in science and simulation.
FP64 stands for “64-bit floating point number” — a method by which a chip stores and computes with decimal numbers. The 64 describes how many bits, i.e. binary digits, are reserved for a single number. The more bits, the more decimal places can be represented, and the smaller the error over long calculations. FP64 is considered the most precise format in widespread everyday use. That’s why it’s also called “double precision” — because it uses twice as many bits as the older FP32.
Precision as a scarce resource
Why should there be different formats at all? More bits cost more: more memory, more power, more time. A chip computing in FP64 needs significantly more resources for the same task than one computing in FP32 or FP16. That’s not a problem as long as precision matters more than speed.
In science, that’s often the case. Small computational errors can accumulate over millions of steps and render a simulation’s result useless. For physical simulations, climate models, or space-flight calculations, FP64 is therefore the standard. An error in the sixth decimal place can there determine success or failure.
How 64 bits describe a number
The 64 bits are divided into three groups. A single bit stores the sign — positive or negative. Eleven bits describe the exponent, i.e. the magnitude of the number. The remaining 52 bits store the actual digits, the so-called mantissa. Together, this allows numbers between roughly 10 to the power of minus 308 and 10 to the power of 308 to be represented, with up to 15 to 16 significant decimal digits.
By comparison: FP32 has only 23 bits for the mantissa and achieves about 7 significant digits. FP16, which is popular in AI applications, manages even fewer. The difference sounds abstract, but becomes concrete when you consider: if you add the number 1 to a million in FP16, the result may still be a million — because the format simply discards the small decimal places. FP64 wouldn’t have this problem.
Modern graphics chips, so-called GPUs, were originally built for game graphics. Low precision is sufficient for that. That’s why many GPUs are optimized for FP32 or FP16 and compute significantly more slowly in FP64 — sometimes at only one-sixteenth of maximum speed. Chips for scientific computing, such as Nvidia's H100, therefore have a separate, stronger FP64 unit.
FP64 in AI chips, supercomputers, and benchmarks
In AI research, FP64 is the exception. Training large language models runs almost always in FP16 or BF16, because speed matters more there than maximum precision. The models learn from vast amounts of data and in any case average out many errors along the way. Small inaccuracies barely matter.
Nevertheless, FP64 regularly shows up in the AI context — as a figure in data sheets. Chip manufacturers state FP64 performance in TFLOPS, i.e. trillions of floating-point operations per second. This figure serves as a comparison point between different chips, even though most AI workloads don’t make use of it at all. So anyone reading chip benchmarks should check whether the impressive TFLOPS figure applies to FP64, FP32, or FP16 — the differences are substantial.
In practice, FP64 is encountered mainly where AI meets classical science: for instance, in weather forecasting, in simulating molecules for drug development, or in training physical models. Here, FP16 isn’t enough, and FP64 isn’t a luxury option but simply a necessity.