FP64

FP64, or double precision, is a 64-bit floating point format with 1 sign bit, 11 exponent bits, and 52 mantissa bits. It stores each value in 8 bytes and delivers about 15 to 17 decimal digits of precision. That's the level scientific simulation and other HPC workloads need to keep rounding error from compounding over long computations.

FP64 range runs to roughly 1.8 times ten to the 308th, far beyond FP32. The catch is throughput. Data center GPUs like the A100 and H100 run FP64 at close to half their FP32 rate. Consumer GeForce cards run FP64 at a fraction of that, often 1/32 or 1/64 of FP32 throughput, because the silicon carries fewer FP64 execution units.

A workload that genuinely needs FP64, computational fluid dynamics or climate modeling for instance, will bottleneck hard on a consumer card regardless of its FP32 or FP16 numbers. Check the FP64 TFLOPS line specifically, not just the FP32 headline figure, before specing a GPU for that kind of work.

Sources

Last verified August 29, 2026.