diff --git a/.pick_status.json b/.pick_status.json
index b6b4403115f..86b8dce9f2c 100644
--- a/.pick_status.json
+++ b/.pick_status.json
@@ -616,7 +616,7 @@
         "description": "winsys/radeon: fix the scratch buffer on gfx6-7",
         "nominated": true,
         "nomination_type": 1,
-        "resolution": 0,
+        "resolution": 1,
         "main_sha": null,
         "because_sha": "1bf39b1f9d115d69aa7b192beb7cde5eea31dffe"
     },
diff --git a/src/gallium/winsys/radeon/drm/radeon_drm_winsys.c b/src/gallium/winsys/radeon/drm/radeon_drm_winsys.c
index ce099dc841e..d748d2a6c7b 100644
--- a/src/gallium/winsys/radeon/drm/radeon_drm_winsys.c
+++ b/src/gallium/winsys/radeon/drm/radeon_drm_winsys.c
@@ -586,6 +586,13 @@ static bool do_winsys_init(struct radeon_drm_winsys *ws)
    ws->info.spi_cu_en = 0xffff;
    ws->info.never_stop_sq_perf_counters = false;
 
+   /* The maximum number of scratch waves. The number is only a function of the number of CUs.
+    * It should be large enough to hold at least 1 threadgroup. Use the minimum per-SA CU count.
+    */
+   const unsigned max_waves_per_tg = 1024 / 64; /* LLVM only supports 1024 threads per block */
+   ws->info.max_scratch_waves = MAX2(32 * ws->info.min_good_cu_per_sa * ws->info.max_sa_per_se *
+                                     ws->info.num_se, max_waves_per_tg);
+
    ws->check_vm = strstr(debug_get_option("R600_DEBUG", ""), "check_vm") != NULL ||
                                                                             strstr(debug_get_option("AMD_DEBUG", ""), "check_vm") != NULL;
    ws->noop_cs = debug_get_bool_option("RADEON_NOOP", false);