diff --git a/test/unit/test_viz.py b/test/unit/test_viz.py
index 7212b56aff..fbfc37e76f 100644
--- a/test/unit/test_viz.py
+++ b/test/unit/test_viz.py
@@ -379,9 +379,9 @@ class TestVizProfiler(unittest.TestCase):
     j = load_profile(prof)
 
     tracks = list(j['layout'])
-    self.assertEqual(tracks[0], 'NV Graph')
-    self.assertEqual(tracks[1], 'NV')
-    self.assertEqual(tracks[2], 'NV:1')
+    self.assertEqual(tracks[0], 'NV')
+    self.assertEqual(tracks[1], 'NV:1')
+    self.assertEqual(tracks[2], 'NV Graph')
 
     nv_events = j['layout']['NV']['events']
     self.assertEqual(nv_events[0]['name'], 'E_25_4n2')
diff --git a/tinygrad/viz/serve.py b/tinygrad/viz/serve.py
index 22941527c4..e5945fbfcb 100755
--- a/tinygrad/viz/serve.py
+++ b/tinygrad/viz/serve.py
@@ -198,7 +198,8 @@ def get_profile(profile:list[ProfileEvent]) -> bytes|None:
     v.sort(key=lambda e:e[0])
     layout[k] = timeline_layout(v, start_ts, scache)
     layout[f"{k} Memory"] = mem_layout(v, start_ts, unwrap(end_ts), peaks, dtype_size, scache)
-  ret = [b"".join([struct.pack("<B", len(k)), k.encode(), v]) for k,v in layout.items() if v is not None]
+  groups = sorted(layout.items(), key=lambda x: '' if len(ss:=x[0].split(" ")) == 1 else ss[1])
+  ret = [b"".join([struct.pack("<B", len(k)), k.encode(), v]) for k,v in groups if v is not None]
   index = json.dumps({"strings":list(scache), "dtypeSize":dtype_size, "markers":[{"ts":int(e.ts-start_ts), **e.arg} for e in markers]}).encode()
   return struct.pack("<IQII", unwrap(end_ts)-start_ts, max(peaks,default=0), len(index), len(ret))+index+b"".join(ret)