diff --git a/gsk/gpu/gskglframe.c b/gsk/gpu/gskglframe.c
index a5f2a84f6f..367e2e1ffa 100644
--- a/gsk/gpu/gskglframe.c
+++ b/gsk/gpu/gskglframe.c
@@ -4,6 +4,7 @@
 
 #include "gskgpuglobalsopprivate.h"
 #include "gskgpuopprivate.h"
+#include "gskgpushaderopprivate.h"
 #include "gskglbufferprivate.h"
 #include "gskgldeviceprivate.h"
 
@@ -66,6 +67,13 @@ static GskGpuBuffer *
 gsk_gl_frame_create_vertex_buffer (GskGpuFrame *frame,
                                    gsize        size)
 {
+  GskGLFrame *self = GSK_GL_FRAME (frame);
+
+  /* We could also reassign them all to the new buffer here?
+   * Is that faster?
+   */
+  g_hash_table_remove_all (self->vaos);
+
   return gsk_gl_buffer_new (GL_ARRAY_BUFFER, size, GL_WRITE_ONLY);
 }
 
@@ -166,11 +174,12 @@ gsk_gl_frame_use_program (GskGLFrame                *self,
   vao = GPOINTER_TO_UINT (g_hash_table_lookup (self->vaos, op_class));
   if (vao)
     {
-      glBindVertexArray(vao);
+      glBindVertexArray (vao);
       return;
     }
-  glGenVertexArrays(1, &vao);
-  glBindVertexArray(vao);
+  glGenVertexArrays (1, &vao);
+  glBindVertexArray (vao);
+  op_class->setup_vao (0);
 
   g_hash_table_insert (self->vaos, (gpointer) op_class, GUINT_TO_POINTER (vao));
 }
diff --git a/gsk/gpu/gskgpurenderer.c b/gsk/gpu/gskgpurenderer.c
index ed7002e14a..8f41e435a4 100644
--- a/gsk/gpu/gskgpurenderer.c
+++ b/gsk/gpu/gskgpurenderer.c
@@ -24,6 +24,8 @@
 static const GdkDebugKey gsk_gpu_optimization_keys[] = {
   { "uber", GSK_GPU_OPTIMIZE_UBER, "Don't use the uber shader" },
   { "clear", GSK_GPU_OPTIMIZE_CLEAR, "Don't vkCmdClearAttachment()/glClear() instead of shaders" },
+
+  { "gl-baseinstance", GSK_GPU_OPTIMIZE_GL_BASE_INSTANCE, "Assume no ARB/EXT_base_instance support" },
 };
 
 typedef struct _GskGpuRendererPrivate GskGpuRendererPrivate;
@@ -32,6 +34,7 @@ struct _GskGpuRendererPrivate
 {
   GskGpuDevice *device;
   GdkDrawContext *context;
+  GskGpuOptimizations optimizations;
 
   GskGpuFrame *frames[GSK_GPU_MAX_FRAMES];
 };
@@ -82,7 +85,7 @@ gsk_gpu_renderer_create_frame (GskGpuRenderer *self)
 
   result = g_object_new (klass->frame_type, NULL);
 
-  gsk_gpu_frame_setup (result, self, priv->device, klass->optimizations);
+  gsk_gpu_frame_setup (result, self, priv->device, priv->optimizations);
 
   return result;
 }
@@ -118,6 +121,7 @@ gsk_gpu_renderer_realize (GskRenderer  *renderer,
 {
   GskGpuRenderer *self = GSK_GPU_RENDERER (renderer);
   GskGpuRendererPrivate *priv = gsk_gpu_renderer_get_instance_private (self);
+  GskGpuOptimizations context_optimizations;
   GdkDisplay *display;
 
   if (surface)
@@ -129,13 +133,15 @@ gsk_gpu_renderer_realize (GskRenderer  *renderer,
   if (priv->device == NULL)
     return FALSE;
 
-  priv->context = GSK_GPU_RENDERER_GET_CLASS (self)->create_context (self, display, surface, error);
+  priv->context = GSK_GPU_RENDERER_GET_CLASS (self)->create_context (self, display, surface, &context_optimizations, error);
   if (priv->context == NULL)
     {
       g_clear_object (&priv->device);
       return FALSE;
     }
 
+  priv->optimizations &= context_optimizations;
+
   return TRUE;
 }
 
@@ -286,6 +292,9 @@ gsk_gpu_renderer_class_init (GskGpuRendererClass *klass)
 static void
 gsk_gpu_renderer_init (GskGpuRenderer *self)
 {
+  GskGpuRendererPrivate *priv = gsk_gpu_renderer_get_instance_private (self);
+
+  priv->optimizations = GSK_GPU_RENDERER_GET_CLASS (self)->optimizations;
 }
 
 GdkDrawContext *
diff --git a/gsk/gpu/gskgpurendererprivate.h b/gsk/gpu/gskgpurendererprivate.h
index 8a408c6fac..8d2c73221a 100644
--- a/gsk/gpu/gskgpurendererprivate.h
+++ b/gsk/gpu/gskgpurendererprivate.h
@@ -23,13 +23,14 @@ struct _GskGpuRendererClass
   GskRendererClass parent_class;
 
   GType frame_type;
-  GskGpuOptimizations optimizations;
+  GskGpuOptimizations optimizations; /* subclasses cannot override this */
 
   GskGpuDevice *        (* get_device)                                  (GdkDisplay             *display,
                                                                          GError                **error);
   GdkDrawContext *      (* create_context)                              (GskGpuRenderer         *self,
                                                                          GdkDisplay             *display,
                                                                          GdkSurface             *surface,
+                                                                         GskGpuOptimizations    *supported,
                                                                          GError                **error);
 
   void                  (* make_current)                                (GskGpuRenderer         *self);
diff --git a/gsk/gpu/gskgpushaderop.c b/gsk/gpu/gskgpushaderop.c
index 7422ae36a7..f61ddd1edf 100644
--- a/gsk/gpu/gskgpushaderop.c
+++ b/gsk/gpu/gskgpushaderop.c
@@ -96,12 +96,23 @@ gsk_gpu_shader_op_gl_command_n (GskGpuOp    *op,
                      gsk_gl_device_get_sampler_id (device, images[i].sampler));
     }
 
-  shader_op_class->setup_vao (self->vertex_offset);
+  if (gsk_gpu_frame_should_optimize (frame, GSK_GPU_OPTIMIZE_GL_BASE_INSTANCE))
+    {
+      glDrawArraysInstancedBaseInstance (GL_TRIANGLES,
+                                         0,
+                                         6 * instance_scale,
+                                         1,
+                                         self->vertex_offset / shader_op_class->vertex_size);
+    }
+  else
+    {
+      shader_op_class->setup_vao (self->vertex_offset);
 
-  glDrawArraysInstanced (GL_TRIANGLES,
-                         0,
-                         6 * instance_scale,
-                         1);
+      glDrawArraysInstanced (GL_TRIANGLES,
+                             0,
+                             6 * instance_scale,
+                             1);
+    }
 
   return op->next;
 }
diff --git a/gsk/gpu/gskgputypesprivate.h b/gsk/gpu/gskgputypesprivate.h
index 0d83ed9bc4..d3e56d7650 100644
--- a/gsk/gpu/gskgputypesprivate.h
+++ b/gsk/gpu/gskgputypesprivate.h
@@ -57,5 +57,7 @@ typedef enum {
 typedef enum {
   GSK_GPU_OPTIMIZE_UBER                 = 1 <<  0,
   GSK_GPU_OPTIMIZE_CLEAR                = 1 <<  1,
+  /* These require hardware support */
+  GSK_GPU_OPTIMIZE_GL_BASE_INSTANCE     = 1 <<  2,
 } GskGpuOptimizations;
 
diff --git a/gsk/gpu/gsknglrenderer.c b/gsk/gpu/gsknglrenderer.c
index 4319fc1061..a0b5478757 100644
--- a/gsk/gpu/gsknglrenderer.c
+++ b/gsk/gpu/gsknglrenderer.c
@@ -8,6 +8,8 @@
 #include "gskglframeprivate.h"
 #include "gskglimageprivate.h"
 
+#include "gdk/gdkglcontextprivate.h"
+
 struct _GskNglRenderer
 {
   GskGpuRenderer parent_instance;
@@ -23,10 +25,11 @@ struct _GskNglRendererClass
 G_DEFINE_TYPE (GskNglRenderer, gsk_ngl_renderer, GSK_TYPE_GPU_RENDERER)
 
 static GdkDrawContext *
-gsk_ngl_renderer_create_context (GskGpuRenderer *renderer,
-                                 GdkDisplay     *display,
-                                 GdkSurface     *surface,
-                                 GError        **error)
+gsk_ngl_renderer_create_context (GskGpuRenderer       *renderer,
+                                 GdkDisplay           *display,
+                                 GdkSurface           *surface,
+                                 GskGpuOptimizations  *supported,
+                                 GError              **error)
 {
   GdkGLContext *context;
 
@@ -47,6 +50,14 @@ gsk_ngl_renderer_create_context (GskGpuRenderer *renderer,
       return NULL;
     }
 
+  gdk_gl_context_make_current (context);
+
+  *supported = -1;
+  if (!gdk_gl_context_check_version (context, "4.2", "9.9") &&
+      !epoxy_has_gl_extension ("GL_EXT_base_instance") &&
+      !epoxy_has_gl_extension ("GL_ARB_base_instance"))
+    *supported &= ~GSK_GPU_OPTIMIZE_GL_BASE_INSTANCE;
+
   return GDK_DRAW_CONTEXT (context);
 }
 
diff --git a/gsk/gpu/gskvulkanrenderer.c b/gsk/gpu/gskvulkanrenderer.c
index afe33d2478..b3fdd1cb38 100644
--- a/gsk/gpu/gskvulkanrenderer.c
+++ b/gsk/gpu/gskvulkanrenderer.c
@@ -73,10 +73,11 @@ gsk_vulkan_renderer_update_images_cb (GdkVulkanContext  *context,
 }
 
 static GdkDrawContext *
-gsk_vulkan_renderer_create_context (GskGpuRenderer *renderer,
-                                    GdkDisplay     *display,
-                                    GdkSurface     *surface,
-                                    GError        **error)
+gsk_vulkan_renderer_create_context (GskGpuRenderer       *renderer,
+                                    GdkDisplay           *display,
+                                    GdkSurface           *surface,
+                                    GskGpuOptimizations  *supported,
+                                    GError              **error)
 {
   GskVulkanRenderer *self = GSK_VULKAN_RENDERER (renderer);
   GdkVulkanContext *context;
@@ -95,6 +96,8 @@ gsk_vulkan_renderer_create_context (GskGpuRenderer *renderer,
                     self);
   gsk_vulkan_renderer_update_images_cb (context, self);
 
+  *supported = -1;
+
   return GDK_DRAW_CONTEXT (context);
 }