Sparsity
performance
Also known as: structured sparsity,dense vs sparse,2:4 sparsity
Definition
Structured sparsity lets a GPU skip half the weights in a 2:4 pattern, doubling theoretical throughput. Vendors usually headline the sparse figure, which is exactly twice the dense one — NVIDIA quotes the B200 at 144 PFLOPS FP4 sparse against 72 dense. Flopper publishes dense throughout.