FPU

Floating Point Unit (FPU) refers to the part of a computer's central processing unit (CPU) responsible for performing operations on floating-point numbers. Floating-point arithmetic enables the representation of a wide dynamic range of values, essential for scientific, engineering, and graphics calculations. Modern FPUs typically implement the IEEE 754 standard, including formats such as binary32 (single precision) and binary64 (double precision), plus extended formats in some implementations.

History and integration: In early microprocessors, floating-point support was provided by a separate coprocessor, such as the 80487. With advancements, FPUs were integrated into the CPU core, and later supplemented by SIMD units (SSE, AVX on x86; NEON on ARM) to accelerate vectorized floating-point operations.

Operations and interface: FPUs perform addition, subtraction, multiplication, division, and square roots, plus transcendental functions via library routines or dedicated instructions. They also handle conversions between integer and floating-point representations and support rounding modes (to nearest, toward zero, toward positive or negative infinity) and exceptions such as overflow, underflow, invalid operation, and inexact results. Compliance with IEEE 754 standard defines formats, rounding, and exceptions.

Architecture: Some FPUs, notably the traditional x87, use a stack-based register model with eight registers. Most modern CPUs use a register-based approach with explicit FP registers and vector units (e.g., SSE/AVX on Intel/AMD, NEON on ARM) that operate on multiple data elements in parallel. This enables higher throughput for scientific and multimedia workloads.

Software and performance: Performance depends on latency and throughput, pipelining, compiler optimizations, and microarchitectural features. OS and language runtimes rely on the FPU for arithmetic; many floating-point functions in standard libraries rely on the FPU. Floating-point computation remains subject to rounding and special cases such as denormal numbers.