diff --git a/Quda/Benchmark_Quda.cpp b/Quda/Benchmark_Quda.cpp
index 67c81bc..11d8c6e 100644
--- a/Quda/Benchmark_Quda.cpp
+++ b/Quda/Benchmark_Quda.cpp
@@ -4,6 +4,7 @@
 #include <cassert>
 #include <chrono>
 #include <color_spinor_field.h>
+#include <communicator_quda.h>
 #include <dirac_quda.h>
 #include <fstream>
 #include <gauge_tools.h>
@@ -51,6 +52,10 @@ template <class F> double bench(F const &f, double target_time, int niter_warmup
   // niter = std::min(1000, niter);
   // printfQuda("during warmup took %f s/iter, deciding on %d iters\n", secs, niter);
 
+  // important: each rank has its own timer, so their measurements can slightly vary. But
+  // 'niter' needs to be consistent (bug took me a couple hours to track down)
+  comm_broadcast_global(&niter, sizeof(niter), 0);
+
   timer.reset(__FUNCTION__, __FILE__, __LINE__);
   timer.start();
   for (int iter = 0; iter < niter; ++iter)
diff --git a/Quda/build-benchmark.sh b/Quda/build-benchmark.sh
index 9a9892c..288f3ac 100755
--- a/Quda/build-benchmark.sh
+++ b/Quda/build-benchmark.sh
@@ -28,5 +28,5 @@ mkdir -p "${PREFIX_DIR}"
 
 LINK_FLAGS="-Wl,-rpath,$QUDA_DIR/lib: $QUDA_DIR/lib/libquda.so $EXTRA_LIBS -lpthread -lmpi"
 
-g++ $BUILD_FLAGS  -I$QUDA_DIR/include   -c -o $BUILD_DIR/Benchmark_Quda.o  $script_dir/Benchmark_Quda.cpp
+g++ $BUILD_FLAGS -I$QUDA_DIR/include/targets/cuda  -I$QUDA_DIR/include   -c -o $BUILD_DIR/Benchmark_Quda.o  $script_dir/Benchmark_Quda.cpp
 g++ -g -O3 $BUILD_DIR/Benchmark_Quda.o -o $PREFIX_DIR/Benchmark_Quda $LINK_FLAGS -lmpi