diff --git a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq3_s.comp b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq3_s.comp
index 5cdf2a89d..42f52b4a1 100644
--- a/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq3_s.comp
+++ b/ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_iq3_s.comp
@@ -7,7 +7,14 @@ layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 
 FLOAT_TYPE temp[NUM_COLS][NUM_ROWS];
 
-void calc_superblock(const uint a_offset, const uint b_offset, const uint ib32, const uint i, const uint num_blocks_per_row, const uint first_row, const uint num_rows) {
+// invocations per superblock. with many columns, 8 invocations need too many
+// registers and spill, so use 16 to halve the per-invocation B working set
+const uint TPB = NUM_COLS <= 4 ? 8 : 16;
+const uint NL  = 32 / TPB; // l steps per invocation
+
+void calc_superblock(const uint a_offset, const uint b_offset, const uint itid, const uint i, const uint num_blocks_per_row, const uint first_row, const uint num_rows) {
+    const uint ib32 = itid / (TPB / 8);
+    const uint l0   = (itid % (TPB / 8)) * NL;
     const uint y_idx = i * QUANT_K + 32 * ib32;
 
     uint ibi = a_offset + first_row * num_blocks_per_row + i;
@@ -16,11 +23,8 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint ib32,
         const uint scale = (data_a[ibi].scales[ib32/2] >> (4 * (ib32 & 1))) & 0xF;
         const float dscale = d * (1 + 2 * scale);
         const uint qh = data_a[ibi].qh[ib32];
-        FLOAT_TYPE sum[NUM_COLS];
-        [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
-            sum[j] = 0.0;
-        }
-        [[unroll]] for (uint l = 0; l < 4; ++l) {
+        [[unroll]] for (uint ll = 0; ll < NL; ++ll) {
+            const uint l = l0 + ll;
             const u8vec2 qs = unpack8(uint32_t(data_a_packed16[ibi].qs[4 * ib32 + l])).xy; // vec4 used due to #12147
             const uint sign = data_a[ibi].signs[4 * ib32 + l];
             const vec4 grid0 = vec4(unpack8(iq3s_grid[qs.x | ((qh << (8 - 2*l)) & 0x100)]));
@@ -30,7 +34,7 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint ib32,
                 const vec4 b0 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 0]);
                 const vec4 b4 = vec4(data_b_v4[(j*p.batch_stride_b + b_offset + y_idx) / 4 + 2*l + 1]);
 
-                sum[j] =
+                const FLOAT_TYPE sum =
                       fma(FLOAT_TYPE(b0.x), FLOAT_TYPE((sign &   1) != 0 ? -grid0.x : grid0.x),
                       fma(FLOAT_TYPE(b0.y), FLOAT_TYPE((sign &   2) != 0 ? -grid0.y : grid0.y),
                       fma(FLOAT_TYPE(b0.z), FLOAT_TYPE((sign &   4) != 0 ? -grid0.z : grid0.z),
@@ -39,12 +43,11 @@ void calc_superblock(const uint a_offset, const uint b_offset, const uint ib32,
                       fma(FLOAT_TYPE(b4.y), FLOAT_TYPE((sign &  32) != 0 ? -grid1.y : grid1.y),
                       fma(FLOAT_TYPE(b4.z), FLOAT_TYPE((sign &  64) != 0 ? -grid1.z : grid1.z),
                       fma(FLOAT_TYPE(b4.w), FLOAT_TYPE((sign & 128) != 0 ? -grid1.w : grid1.w),
-                      sum[j]))))))));
+                      FLOAT_TYPE(0.0)))))))));
+
+                temp[j][n] = fma(dscale, sum, temp[j][n]);
             }
         }
-        [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
-            temp[j][n] = fma(dscale, sum[j], temp[j][n]);
-        }
         ibi += num_blocks_per_row;
     }
 }
@@ -55,11 +58,11 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
 
     const uint num_blocks_per_row = p.ncols / QUANT_K;
 
-    // 8 threads are used to process each block
-    const uint blocks_per_wg = gl_WorkGroupSize.x/8;
+    // TPB invocations are used to process each block
+    const uint blocks_per_wg = gl_WorkGroupSize.x/TPB;
     const uint tid = gl_LocalInvocationID.x;
-    const uint itid = tid % 8;  // 0...7
-    const uint ix = tid / 8;
+    const uint itid = tid % TPB;
+    const uint ix = tid / TPB;
 
     [[unroll]] for (uint j = 0; j < NUM_COLS; ++j) {
         [[unroll]] for (uint i = 0; i < NUM_ROWS; ++i) {
diff --git a/tests/test-backend-ops.cpp b/tests/test-backend-ops.cpp
index dc28aba2d..4d9808562 100644
--- a/tests/test-backend-ops.cpp
+++ b/tests/test-backend-ops.cpp
@@ -9395,6 +9395,10 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
             //test_cases.emplace_back(new test_mul_mat(type_a,    GGML_TYPE_F32, 18,  i, 32*256, { 1,  1}, {8, 1}));
             //test_cases.emplace_back(new test_mul_mat(type_a,    GGML_TYPE_F32, 19,  i, 33*256, { 1,  1}, {1, 1}));
         }
+        // mat-vec shaders split k across lanes and loop over the blocks in strides. k must be
+        // long enough that the loop wraps, else the stride is never exercised
+        test_cases.emplace_back(new test_mul_mat(type_a,    GGML_TYPE_F32, 16,  1, 16*256, { 1,  1}, {1, 1}));
+        test_cases.emplace_back(new test_mul_mat(type_a,    GGML_TYPE_F32, 16,  8, 16*256, { 1,  1}, {1, 1}));
     }
 
     test_cases.emplace_back(new test_mul_mat(GGML_TYPE_Q4_0, GGML_TYPE_F32, 2880, 32, 2880, {1, 1}, {1, 1}));