Concept融合乘加 / Fused Multiply-Add / FMA

Fused Multiply-Add (FMA) Instruction

A GPU instruction type that performs a multiply-accumulate operation (a×b+c) in a single instruction. FMA units are the most numerous and highest-throughput arithmetic resources on a GPU, and polynomial evaluation can be fully unrolled into a sequence of FMA instructions.

Timeline (last 90 days)

Sep 16

A single FMA instruction may require 4 to 5 clock cycles to produce a result, but the processor can issue multiple FMA instructions per cycle.

Unverified50%

All Facts (1)

Unverified

A single FMA instruction may require 4 to 5 clock cycles to produce a result, but the processor can issue multiple FMA instructions per cycle.

50%

Source Articles

Polynomial Approximation Accelerates LLM Transcendental Functions: Up to 8% Speedup on GB200Oct 4