diff --git a/docs/envvars.rst b/docs/envvars.rst
index be9096505eddfea82ec5b54df2a5a52741255071..0a0a8e5877abbf17727cfe06489a0ef4e7763196 100644
--- a/docs/envvars.rst
+++ b/docs/envvars.rst
@@ -657,6 +657,8 @@ RADV driver environment variables
       disable memory shaders cache
    ``nongg``
       disable NGG for GFX10+
+   ``nonggc``
+      disable NGG culling on GPUs where it's enabled by default (GFX10.3+ only).
    ``nooutoforder``
       disable out-of-order rasterization
    ``notccompatcmask``
@@ -712,7 +714,7 @@ RADV driver environment variables
    ``pswave32``
       enable wave32 for pixel shaders (GFX10+)
    ``nggc``
-      enable NGG culling on GFX10+ GPUs.
+      enable NGG culling on GPUs where it's not enabled by default (GFX10.1 only).
    ``rt``
       enable rt extensions whose implementation is still experimental.
    ``sam``
diff --git a/docs/relnotes/new_features.txt b/docs/relnotes/new_features.txt
index 3f0d07e81af51dc020f16d369d4cfc9408b439fb..cb5e9be7ec12b833378bb1dc7b929640b010a351 100644
--- a/docs/relnotes/new_features.txt
+++ b/docs/relnotes/new_features.txt
@@ -16,3 +16,4 @@ VK_KHR_shader_subgroup_extended_types on lavapipe
 VK_KHR_spirv_1_4 on lavapipe
 Experimental raytracing support on RADV
 VK_KHR_synchronization2 on Intel
+NGG shader based culling is now enabled by default on GFX10.3 on RADV.
\ No newline at end of file
diff --git a/src/amd/vulkan/radv_cmd_buffer.c b/src/amd/vulkan/radv_cmd_buffer.c
index 7027487b8e8d3a79a3f3b7576feaa0d8aaa4e9ad..ddc73bcd8a82aeb82a798d99009f4ed562bc3efa 100644
--- a/src/amd/vulkan/radv_cmd_buffer.c
+++ b/src/amd/vulkan/radv_cmd_buffer.c
@@ -5927,7 +5927,7 @@ radv_emit_all_graphics_states(struct radv_cmd_buffer *cmd_buffer, const struct r
        cmd_buffer->state.emitted_pipeline != cmd_buffer->state.pipeline)
       radv_emit_rbplus_state(cmd_buffer);
 
-   if ((cmd_buffer->device->instance->perftest_flags & RADV_PERFTEST_NGGC) &&
+   if (cmd_buffer->device->physical_device->use_ngg_culling &&
        cmd_buffer->state.pipeline->graphics.is_ngg)
       radv_emit_ngg_culling_state(cmd_buffer, info);
 
diff --git a/src/amd/vulkan/radv_debug.h b/src/amd/vulkan/radv_debug.h
index 19dc22751461ef4b69c50ee4b8cc6fb475eab4f3..0bfdd4889ceb105179d91c55c03e8e48d7845690 100644
--- a/src/amd/vulkan/radv_debug.h
+++ b/src/amd/vulkan/radv_debug.h
@@ -62,6 +62,7 @@ enum {
    RADV_DEBUG_NO_TC_COMPAT_CMASK = 1ull << 31,
    RADV_DEBUG_NO_VRS_FLAT_SHADING = 1ull << 32,
    RADV_DEBUG_NO_ATOC_DITHERING = 1ull << 33,
+   RADV_DEBUG_NO_NGGC = 1ull << 34,
 };
 
 enum {
diff --git a/src/amd/vulkan/radv_device.c b/src/amd/vulkan/radv_device.c
index a1ff738fa1f29aa4db97085e2f6f0d066306f8fe..362445be5aa41655a8b323c9d661b6f85de05f4d 100644
--- a/src/amd/vulkan/radv_device.c
+++ b/src/amd/vulkan/radv_device.c
@@ -705,6 +705,13 @@ radv_physical_device_try_create(struct radv_instance *instance, drmDevicePtr drm
                      device->rad_info.family != CHIP_NAVI14 &&
                      !(device->instance->debug_flags & RADV_DEBUG_NO_NGG);
 
+   device->use_ngg_culling =
+      device->use_ngg &&
+      device->rad_info.max_render_backends > 1 &&
+      (device->rad_info.chip_class >= GFX10_3 ||
+       (device->instance->perftest_flags & RADV_PERFTEST_NGGC)) &&
+      !(device->instance->debug_flags & RADV_DEBUG_NO_NGGC);
+
    device->use_ngg_streamout = false;
 
    /* Determine the number of threads per wave for all stages. */
@@ -841,6 +848,7 @@ static const struct debug_control radv_debug_options[] = {
    {"notccompatcmask", RADV_DEBUG_NO_TC_COMPAT_CMASK},
    {"novrsflatshading", RADV_DEBUG_NO_VRS_FLAT_SHADING},
    {"noatocdithering", RADV_DEBUG_NO_ATOC_DITHERING},
+   {"nonggc", RADV_DEBUG_NO_NGGC},
    {NULL, 0}};
 
 const char *
diff --git a/src/amd/vulkan/radv_pipeline.c b/src/amd/vulkan/radv_pipeline.c
index 69510fdbf537487797cb657d5671ac139ad7d275..581a77fa808d86e75c25d0acf62c308b838056b5 100644
--- a/src/amd/vulkan/radv_pipeline.c
+++ b/src/amd/vulkan/radv_pipeline.c
@@ -217,8 +217,8 @@ radv_get_hash_flags(const struct radv_device *device, bool stats)
 {
    uint32_t hash_flags = 0;
 
-   if (device->instance->perftest_flags & RADV_PERFTEST_NGGC)
-      hash_flags |= RADV_HASH_SHADER_FORCE_NGG_CULLING;
+   if (device->physical_device->use_ngg_culling)
+      hash_flags |= RADV_HASH_SHADER_USE_NGG_CULLING;
    if (device->instance->perftest_flags & RADV_PERFTEST_FORCE_EMULATE_RT)
       hash_flags |= RADV_HASH_SHADER_FORCE_EMULATE_RT;
    if (device->physical_device->cs_wave_size == 32)
diff --git a/src/amd/vulkan/radv_private.h b/src/amd/vulkan/radv_private.h
index 678d9d553dd57cf09e6f08137f428d9a8927562d..f49d59ca8df50f8336011ddeb5f151d38e3b6b1a 100644
--- a/src/amd/vulkan/radv_private.h
+++ b/src/amd/vulkan/radv_private.h
@@ -262,6 +262,9 @@ struct radv_physical_device {
    /* Whether to enable NGG. */
    bool use_ngg;
 
+   /* Whether to enable NGG culling. */
+   bool use_ngg_culling;
+
    /* Whether to enable NGG streamout. */
    bool use_ngg_streamout;
 
@@ -1649,7 +1652,7 @@ struct radv_event {
 #define RADV_HASH_SHADER_GE_WAVE32         (1 << 3)
 #define RADV_HASH_SHADER_LLVM              (1 << 4)
 #define RADV_HASH_SHADER_KEEP_STATISTICS   (1 << 8)
-#define RADV_HASH_SHADER_FORCE_NGG_CULLING (1 << 13)
+#define RADV_HASH_SHADER_USE_NGG_CULLING   (1 << 13)
 #define RADV_HASH_SHADER_ROBUST_BUFFER_ACCESS (1 << 14)
 #define RADV_HASH_SHADER_ROBUST_BUFFER_ACCESS2 (1 << 15)
 #define RADV_HASH_SHADER_FORCE_EMULATE_RT (1 << 16)
diff --git a/src/amd/vulkan/radv_shader.c b/src/amd/vulkan/radv_shader.c
index 6c37a72c44b3a9cdfa9d8347fdd3152fa58c1210..3069f2d02db1de79bc1ec5db976e9da48f0b6c59 100644
--- a/src/amd/vulkan/radv_shader.c
+++ b/src/amd/vulkan/radv_shader.c
@@ -899,10 +899,7 @@ radv_consider_culling(struct radv_device *device, struct nir_shader *nir,
    if (nir->info.outputs_written & (VARYING_BIT_VIEWPORT | VARYING_BIT_VIEWPORT_MASK))
       return false;
 
-   /* TODO: enable by default on GFX10.3 when we're confident about performance. */
-   bool culling_enabled = device->instance->perftest_flags & RADV_PERFTEST_NGGC;
-
-   if (!culling_enabled)
+   if (!device->physical_device->use_ngg_culling)
       return false;
 
    /* Shader based culling efficiency can depend on PS throughput.
@@ -912,9 +909,7 @@ radv_consider_culling(struct radv_device *device, struct nir_shader *nir,
    unsigned max_render_backends = device->physical_device->rad_info.max_render_backends;
    unsigned max_se = device->physical_device->rad_info.max_se;
 
-   if (max_render_backends < 2)
-      return false; /* Don't use NGG culling on 1 RB chips. */
-   else if (max_render_backends / max_se == 4)
+   if (max_render_backends / max_se == 4)
       max_ps_params = 6; /* Sienna Cichlid and other GFX10.3 dGPUs. */
    else
       max_ps_params = 4; /* Navi 1x. */