The paper reports that its query-sparse Triton kernel reaches a 1.99x forward-plus-backward speedup over the dense Triton baseline at 64k sequence length and a 0.5 attention budget.

来源：paper-fixed:Section 4.3 and Figure 2, p. 8