FlushToZero

Flush-to-zero (FTZ) is a mode used by floating-point units in computer processors to handle underflow. When FTZ is enabled, any result that would be denormal (subnormal) is replaced with zero instead of a small, nonzero subnormal value. This behavior is supported in various architectures and is often described together with the Denormals Are Zero (DAZ) mode.

FTZ is motivated by performance: subnormal numbers require special handling and can be much slower on many hardware implementations. By flushing underflows to zero, throughput can improve in numerical kernels, graphics, and signal processing. The trade-off is a potential loss of precision near the smallest magnitudes; some numerical results might differ from strict IEEE 754 behavior when underflow occurs.

Implementation and control: In x86 processors with SSE/AVX, FTZ and DAZ are controlled by bits in the MXCSR control and status register. The exact effect depends on whether FTZ is enabled for results and whether DAZ is enabled for denormal inputs. Other architectures, such as ARM's NEON/VFP, provide analogous options via their floating-point control registers.

Software implications: Many numerical libraries enable FTZ by default to maximize performance, but applications requiring exact subnormal handling can disable it at runtime or compile time where supported. When porting code across platforms with different FTZ/DAZ behavior, numerical results near underflow should be validated.

See also: Denormal number; IEEE 754; Floating-point unit.