FP64
The double-precision floating-point format, a high-precision numeric type required for classical scientific simulation such as climate and fusion modeling. In AI-first supercomputers it is often replaced by low-precision matrix operations or emulated.
Timeline (last 90 days)
FP64 (double-precision floating point) is the 64-bit floating-point operation format in the IEEE 754 standard, providing about 15-17 significant decimal digits of precision.
GPU throughput on FP32 is usually far higher than on FP64, and the FP64 capability of consumer-grade GPUs is often substantially cut down.
FP64 has a machine precision of about 2.2×10⁻¹⁶, whereas FP32 has a machine precision of only about 1.2×10⁻⁷.
On GPUs, FP64 theoretical throughput is typically only 1/2 to 1/32 that of FP32 (depending on architecture) and requires more registers and bandwidth, so optimizing FP64 GEMM is more challenging than lower precision.
NVIDIA disclosed on its developer blog that the cuBLAS library has added a batch of optimizations targeting FP64, aimed at accelerating scientific simulation.
The new algorithm paths may differ subtly from old implementations in floating-point rounding order, and simulations sensitive to results need regression comparisons to verify numerical consistency.
Molecular dynamics software typically uses FP64 as the default precision, because error accumulation in force-field integration directly affects energy conservation and trajectory stability.
FP64 (double-precision floating point) means representing a floating-point number in 64 binary bits, of which 52 bits are for the mantissa, 11 bits for the exponent and 1 bit for the sign.
All Facts (8)
FP64 (double-precision floating point) is the 64-bit floating-point operation format in the IEEE 754 standard, providing about 15-17 significant decimal digits of precision.
50%UnverifiedGPU throughput on FP32 is usually far higher than on FP64, and the FP64 capability of consumer-grade GPUs is often substantially cut down.
50%UnverifiedFP64 has a machine precision of about 2.2×10⁻¹⁶, whereas FP32 has a machine precision of only about 1.2×10⁻⁷.
50%UnverifiedOn GPUs, FP64 theoretical throughput is typically only 1/2 to 1/32 that of FP32 (depending on architecture) and requires more registers and bandwidth, so optimizing FP64 GEMM is more challenging than lower precision.
50%UnverifiedThe new algorithm paths may differ subtly from old implementations in floating-point rounding order, and simulations sensitive to results need regression comparisons to verify numerical consistency.
50%UnverifiedMolecular dynamics software typically uses FP64 as the default precision, because error accumulation in force-field integration directly affects energy conservation and trajectory stability.
50%UnverifiedFP64 (double-precision floating point) means representing a floating-point number in 64 binary bits, of which 52 bits are for the mantissa, 11 bits for the exponent and 1 bit for the sign.
50%UnverifiedNVIDIA disclosed on its developer blog that the cuBLAS library has added a batch of optimizations targeting FP64, aimed at accelerating scientific simulation.
50%