diff --git a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_mx_moe_v3.hpp b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_mx_moe_v3.hpp
index df5b709222..bb4286b3f5 100644
--- a/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_mx_moe_v3.hpp
+++ b/include/ck/tensor_operation/gpu/block/blockwise_gemm_pipeline_xdlops_mx_moe_v3.hpp
@@ -218,9 +218,9 @@ struct BlockwiseGemmXdlops_pipeline_mx_moe_nbs_v3<BlockGemmPipelineScheduler::In
             HotLoopInstList::B_LDS_Read_Width * sizeof(BDataType) == 16 ? 8 : 4;
 
         constexpr auto ds_read_a_mfma_rate =
-            (mfma_cycle - 4 + 2 * ds_read_a_issue_cycle - 1) / (2 * ds_read_a_issue_cycle);
+            (mfma_cycle - 8 + 2 * ds_read_a_issue_cycle - 1) / (2 * ds_read_a_issue_cycle);
         constexpr auto ds_read_b_mfma_rate =
-            (mfma_cycle - 4 + 2 * ds_read_b_issue_cycle - 1) / (2 * ds_read_b_issue_cycle);
+            (mfma_cycle - 8 + 2 * ds_read_b_issue_cycle - 1) / (2 * ds_read_b_issue_cycle);
 
         constexpr auto num_dsread_a_mfma =
             (num_ds_read_inst_a + ds_read_a_mfma_rate - 1) / ds_read_a_mfma_rate;