GPU Glossary
Device Hardware·1 min read

Special Function UnitSFU

Hardware units in a Streaming Multiprocessor that accelerate certain arithmetic operations, notably transcendental functions like exp, sin and cos.

Special Function Units (SFUs) are units in Streaming Multiprocessors (SMs) that accelerate certain arithmetic operations. For neural network workloads, the notable ones are transcendental functions like exp, sin and cos.

The internal architecture of an H100 SM. Special Function Units are shown in pink, along with the Load/Store Units. Modified from NVIDIA's H100 white paper.
The internal architecture of an H100 SM. Special Function Units are shown in pink, along with the Load/Store Units. Modified from NVIDIA's H100 white paper.

Streaming Assembler (SASS) instructions for the SFUs generally begin with MUFU, as in MUFU.SQRT and MUFU.EX2. This Godbolt example shows MUFU.EX2 being used to implement the expf intrinsic in CUDA C++.