diff --git a/dev/cuda/gelu_forward.cu b/dev/cuda/gelu_forward.cu
index a8edf31..464f034 100644
--- a/dev/cuda/gelu_forward.cu
+++ b/dev/cuda/gelu_forward.cu
@@ -182,7 +182,7 @@ int main(int argc, const char **argv) {
         // napkin math: estimate the memory bandwidth achieved
         // for each (B,T,C) output element, we do 1 read and 1 write, 4 bytes each
         // and e.g. A100 40GB PCIe is advertised at 1,555GB/s
-        long memory_ops = B * T * C * 2 * 4;
+        long memory_ops = B * T * C * 2 * (int)sizeof(floatX);
         float memory_bandwidth = memory_ops / elapsed_time / 1e6;
 
         printf("block_size %4d | time %.4f ms | bandwidth %.2f GB/s\n", block_size, elapsed_time, memory_bandwidth);