diff --git a/ggml/src/ggml-cuda/mmq.cu b/ggml/src/ggml-cuda/mmq.cu
index 9beff0d9b7..9b6038adff 100644
--- a/ggml/src/ggml-cuda/mmq.cu
+++ b/ggml/src/ggml-cuda/mmq.cu
@@ -171,7 +171,7 @@ void ggml_cuda_mul_mat_q(
             ne00, ne01, ne1, s01, ne11, s1,
             ne02, ne12, s02, s12, s2,
             ne03, ne13, s03, s13, s3,
-            ne1};
+            ne1, ne1};
         ggml_cuda_mul_mat_q_switch_type(ctx, args, stream);
         return;
     }
@@ -244,6 +244,13 @@ void ggml_cuda_mul_mat_q(
                                          ne11 * ne10_padded * sizeof(block_q8_1) / (QK8_1 * sizeof(int));
     const int64_t s13 = ne12*s12;
 
+    // Each expert only sees ne12*n_expert_used/ne02 tokens on average.
+    // On RDNA3 and RDNA4 it is faster to pick the tile size against this value instead of ne12.
+    int64_t ncols_opt = ne12;
+    if (GGML_CUDA_CC_IS_RDNA3_0(cc) || GGML_CUDA_CC_IS_RDNA4(cc)) {
+        ncols_opt = (ne12*n_expert_used + ne02 - 1) / ne02;
+    }
+
     // Note that ne02 is used instead of ne12 because the number of y channels determines the z dimension of the CUDA grid.
     const mmq_args args = {
         src0_d, src0->type, (const int *) src1_q8_1.get(), ids_dst.get(), expert_bounds.get(), dst_d,
@@ -251,7 +258,7 @@ void ggml_cuda_mul_mat_q(
         ne00, ne01, ne_get_rows, s01, ne_get_rows, s1,
         ne02, ne02, s02, s12, s2,
         ne03, ne13, s03, s13, s3,
-        ne12};
+        ne12, ncols_opt};
 
     ggml_cuda_mul_mat_q_switch_type(ctx, args, stream);
 }
diff --git a/ggml/src/ggml-cuda/mmq.cuh b/ggml/src/ggml-cuda/mmq.cuh
index b4a747720f..24afedd143 100644
--- a/ggml/src/ggml-cuda/mmq.cuh
+++ b/ggml/src/ggml-cuda/mmq.cuh
@@ -1376,6 +1376,7 @@ struct mmq_args {
     int64_t nchannels_x; int64_t nchannels_y; int64_t stride_channel_x; int64_t stride_channel_y; int64_t stride_channel_dst;
     int64_t nsamples_x; int64_t nsamples_y; int64_t stride_sample_x; int64_t stride_sample_y; int64_t stride_sample_dst;
     int64_t ncols_max;
+    int64_t ncols_opt; // value to optimize the tile size against, launch grid still uses ncols_max
 };
 
 static size_t mmq_get_nbytes_shared(const ggml_cuda_mmq_config & config, const int cc) {
@@ -1486,7 +1487,7 @@ void mul_mat_q_switch_J(ggml_backend_cuda_context & ctx, const mmq_args & args,
             continue;
         }
 
-        const int ntiles_x = (args.ncols_max + config.J - 1) / config.J;
+        const int ntiles_x = (args.ncols_opt + config.J - 1) / config.J;
 
         if (ntiles_x < ntiles_J_best) {
             J_best = J;