diff --git a/src/freedreno/vulkan/tu_cmd_buffer.cc b/src/freedreno/vulkan/tu_cmd_buffer.cc
index db8e77255da..6af80d78cac 100644
--- a/src/freedreno/vulkan/tu_cmd_buffer.cc
+++ b/src/freedreno/vulkan/tu_cmd_buffer.cc
@@ -1310,7 +1310,7 @@ use_hw_binning(struct tu_cmd_buffer *cmd)
       return true;
    }
 
-   return vsc->binning;
+   return vsc->binning_possible && vsc->binning_useful;
 }
 
 static bool
@@ -1373,8 +1373,16 @@ use_sysmem_rendering(struct tu_cmd_buffer *cmd,
       return true;
    }
 
-   if (TU_DEBUG(GMEM))
+   if (TU_DEBUG(GMEM)) {
+      cmd->state.rp.gmem_disable_reason = "TU_DEBUG(GMEM)";
       return false;
+   }
+
+   /* This is a case where it's better to avoid GMEM, too many tiles but no HW binning possible. */
+   if (!vsc->binning_possible && vsc->binning_useful) {
+      cmd->state.rp.gmem_disable_reason = "Too many tiles and HW binning is not possible";
+      return true;
+   }
 
    bool use_sysmem = cmd->device->autotune->get_optimal_mode(cmd, rp_ctx) == tu_autotune::render_mode::SYSMEM;
    if (use_sysmem)
@@ -6413,7 +6421,7 @@ tu_emit_subpass_begin_gmem(struct tu_cmd_buffer *cmd, struct tu_resolve_group *r
     * (perf queries), then we can't do this optimization since the
     * start-of-the-CS geometry condition will have been overwritten.
     */
-   bool cond_load_allowed = vsc->binning &&
+   bool cond_load_allowed = vsc->binning_possible &&
                             cmd->state.pass->has_cond_load_store &&
                             !cmd->state.rp.draw_cs_writes_to_cond_pred;
 
diff --git a/src/freedreno/vulkan/tu_device.h b/src/freedreno/vulkan/tu_device.h
index 0ac763ef847..979adfbee68 100644
--- a/src/freedreno/vulkan/tu_device.h
+++ b/src/freedreno/vulkan/tu_device.h
@@ -564,8 +564,11 @@ struct tu_vsc_config {
    /* Whether binning could be used for gmem rendering using this framebuffer. */
    bool binning_possible;
 
-   /* Whether binning should be used for gmem rendering using this framebuffer. */
-   bool binning;
+   /* Whether binning is useful for GMEM rendering performance using this framebuffer. This is independent of whether
+    * binning is possible, and is determined by the tile count. Not binning when it's useful would be a performance
+    * hazard, and GMEM rendering should be avoided in the case where it's useful to bin but not possible to do so.
+    */
+   bool binning_useful;
 
    /* pipe register values */
    uint32_t pipe_config[MAX_VSC_PIPES];
diff --git a/src/freedreno/vulkan/tu_util.cc b/src/freedreno/vulkan/tu_util.cc
index e19d43bb8a9..0ebceae8998 100644
--- a/src/freedreno/vulkan/tu_util.cc
+++ b/src/freedreno/vulkan/tu_util.cc
@@ -460,16 +460,12 @@ tu_tiling_config_update_pipes(struct tu_vsc_config *vsc,
 static void
 tu_tiling_config_update_binning(struct tu_vsc_config *vsc, const struct tu_device *device)
 {
-   if (vsc->binning_possible) {
-      vsc->binning = (vsc->tile_count.width * vsc->tile_count.height) > 2;
+   vsc->binning_useful = (vsc->tile_count.width * vsc->tile_count.height) > 2;
 
-      if (TU_DEBUG(FORCEBIN))
-         vsc->binning = true;
-      if (TU_DEBUG(NOBIN))
-         vsc->binning = false;
-   } else {
-      vsc->binning = false;
-   }
+   if (TU_DEBUG(FORCEBIN))
+      vsc->binning_useful = true;
+   if (TU_DEBUG(NOBIN))
+      vsc->binning_useful = false;
 }
 
 void