diff --git a/tests/unit/test_tmem_4warp_read.cu b/tests/unit/test_tmem_4warp_read.cu
index a3f26922..e32ee1d4 100644
--- a/tests/unit/test_tmem_4warp_read.cu
+++ b/tests/unit/test_tmem_4warp_read.cu
@@ -153,7 +153,8 @@ int main() {
     float* d_r;
     cudaMalloc(&d_r, 32 * sizeof(float));
     cudaMemset(d_r, 0, 32 * sizeof(float));
-    test_mma_4warp_read<<<1, 192, 4096>>>(d_r);
+    int smem = 256 + 128 + 128*16*2*2 + 256;  // sbuf + align + sA + sB + slack
+    test_mma_4warp_read<<<1, 192, smem>>>(d_r);
 
     cudaError_t err = cudaDeviceSynchronize();
     if (err != cudaSuccess) {