diff --git a/.github/workflows/good_defines.txt b/.github/workflows/good_defines.txt
index c0f6c73287..cdfaba0e22 100644
--- a/.github/workflows/good_defines.txt
+++ b/.github/workflows/good_defines.txt
@@ -1,3 +1,4 @@
+ALLOW_GPU_PRINTF
 AMREX_DEBUG
 AMREX_PARTICLES
 AMREX_SPACEDIM
diff --git a/Docs/source/mpi_plus_x.rst b/Docs/source/mpi_plus_x.rst
index 7d88e7be77..33df1d1224 100644
--- a/Docs/source/mpi_plus_x.rst
+++ b/Docs/source/mpi_plus_x.rst
@@ -130,6 +130,37 @@ To enable this, compile with::
   USE_HIP = TRUE
 
 
+Printing Warnings from GPU Kernels
+==================================
+
+.. index:: USE_GPU_PRINTF
+
+Castro will output warnings if several assumptions are violated (often
+triggering a retry in the process).  On GPUs, printing from a kernel
+(using ``printf()``) can increase the number of registers a kernel needs,
+causing performance problems.  As a result, warnings are disabled by
+wrapping them in ``#ifndef AMREX_USE_GPU``.
+
+However, for debugging GPU runs, sometimes we want to see these
+warnings.  The build option ``USE_GPU_PRINTF=TRUE`` will enable these
+(by setting the preprocessor flag ``ALLOW_GPU_PRINTF``).
+
+.. note::
+
+   Not every warning has been enabled for GPUs.
+
+.. tip::
+
+   On AMD architectures, it seems necessary to use unbuffered I/O.  This
+   can be accomplished in the job submission script (for SLURM) by doing
+
+   ::
+
+      srun -u ./Castro...
+
+
+
+
 Working at Supercomputing Centers
 =================================
 
diff --git a/Exec/Make.Castro b/Exec/Make.Castro
index 910a577563..5b1575820b 100644
--- a/Exec/Make.Castro
+++ b/Exec/Make.Castro
@@ -136,6 +136,10 @@ ifeq ($(USE_GPU),TRUE)
   endif
 endif
 
+ifeq ($(USE_GPU_PRINTF),TRUE)
+  DEFINES += -DALLOW_GPU_PRINTF
+endif
+
 CASTRO_AUTO_SOURCE_DIR := $(TmpBuildDir)/castro_sources/$(optionsSuffix).EXE
 
 
diff --git a/Source/driver/Castro.cpp b/Source/driver/Castro.cpp
index e79cf0574e..3108b7794f 100644
--- a/Source/driver/Castro.cpp
+++ b/Source/driver/Castro.cpp
@@ -3300,6 +3300,9 @@ Castro::check_for_negative_density ()
                         std::cout << "Invalid X[" << n << "] = " << X << " in zone "
                                   << i << ", " << j << ", " << k
                                   << " with density = " << rho << "\n";
+#elif defined(ALLOW_GPU_PRINTF)
+                        AMREX_DEVICE_PRINTF("Invalid X[%d] = %g in zone (%d,%d,%d) with density = %g\n",
+                                            n, X, i, j, k, rho);
 #endif
                         X_check_failed = 1;
                     }
@@ -3310,6 +3313,10 @@ Castro::check_for_negative_density ()
             return {rho_check_failed, X_check_failed};
         });
 
+#ifdef ALLOW_GPU_PRINTF
+        std::fflush(nullptr);
+#endif
+
     }
 
     ReduceTuple hv = reduce_data.value();