diff --git a/test/test_linearizer.py b/test/test_linearizer.py
index 5f8da35830..8faf467dbb 100644
--- a/test/test_linearizer.py
+++ b/test/test_linearizer.py
@@ -41,7 +41,7 @@ def helper_tc_allclose(n:int, m:int, k:int, dtype_in:DType, dtype_out:DType, axi
   assert len([x for x in k.applied_opts if x.op is OptOps.TC]) == 1, "tensor core opt not included"
   np_c = np_a @ np_b
   if dtype_out == dtypes.half: tc_atol, tc_rtol = 1e-2, 1e-3
-  elif dtype_in == dtypes.bfloat16: tc_atol, tc_rtol = 1e-2, 3e-3
+  elif dtype_out == dtypes.bfloat16: tc_atol, tc_rtol = 1e-2, 1e-2
   else: tc_atol, tc_rtol = 5e-3, 1e-4
   np.testing.assert_allclose(np_c, out, atol=tc_atol, rtol=tc_rtol)
 
@@ -1036,14 +1036,17 @@ class TestLinearizer(unittest.TestCase):
   @unittest.skipUnless(Device[Device.DEFAULT].renderer.tensor_cores, "test requires tensor cores")
   def test_tensor_cores(self):
     for tc in Device[Device.DEFAULT].renderer.tensor_cores:
-      if (getenv("EMULATE_CUDA") or getenv("EMULATE_INTEL")) and (tc.dtype_in == dtypes.bfloat16 or tc.dtype_out == dtypes.bfloat16): continue
+      if (getenv("EMULATE_CUDA") or getenv("EMULATE_INTEL") or getenv("EMULATE_METAL")) and \
+        (tc.dtype_in == dtypes.bfloat16 or tc.dtype_out == dtypes.bfloat16): continue
+      if CI and Device.DEFAULT == "METAL" and (tc.dtype_in == dtypes.bfloat16 or tc.dtype_out == dtypes.bfloat16): continue
       # for AMX, tc.dims[2] == 1 so reduceop is None thus tensor_cores are not triggered
       helper_tc_allclose(tc.dims[0], tc.dims[1], 2 if AMX else tc.dims[2], tc.dtype_in, tc.dtype_out, axis=0, tc_opt=0)
 
   @unittest.skipUnless(Device[Device.DEFAULT].renderer.tensor_cores, "test requires tensor cores")
   def test_tensor_cores_padded(self):
     for tc in Device[Device.DEFAULT].renderer.tensor_cores:
-      if getenv("EMULATE_CUDA") and (tc.dtype_in == dtypes.bfloat16 or tc.dtype_out == dtypes.bfloat16): continue
+      if (getenv("EMULATE_CUDA") or getenv("EMULATE_METAL")) and (tc.dtype_in == dtypes.bfloat16 or tc.dtype_out == dtypes.bfloat16): continue
+      if CI and Device.DEFAULT == "METAL" and (tc.dtype_in == dtypes.bfloat16 or tc.dtype_out == dtypes.bfloat16): continue
       pad = 1
 
       # check that TC is triggered for TC_OPT=2
diff --git a/tinygrad/renderer/cstyle.py b/tinygrad/renderer/cstyle.py
index 331b2cbbcf..3510f11446 100644
--- a/tinygrad/renderer/cstyle.py
+++ b/tinygrad/renderer/cstyle.py
@@ -250,7 +250,8 @@ class MetalRenderer(CStyleLanguage):
   shared_max = 32768
   tensor_cores = [TensorCore(dims=(8,8,8),threads=[(0,2),(1,4),(0,2),(1,2)],expanded_shape=(2,2,2,2),upcast_axes=([(1,2)],[(1,2)],[(1,2)]),
     st1_pattern=(((1,1),(0,1),(1,0),(0,3)),((0,0),(0,2),(1,3),(1,2))),st2_pattern=(((0,0),(1,1),(1,2),(0,2),(1,0)),((0,1),(0,3),(1,3))),
-    dtype_in=di,dtype_out=do,reduce_axes=[(0,8)]) for di,do in [(dtypes.float,dtypes.float),(dtypes.half,dtypes.float),(dtypes.half,dtypes.half)]]
+    dtype_in=di,dtype_out=do,reduce_axes=[(0,8)]) for di,do in [(dtypes.float,dtypes.float),(dtypes.half,dtypes.float),(dtypes.half,dtypes.half),
+                                                                (dtypes.bfloat16,dtypes.float),(dtypes.bfloat16,dtypes.bfloat16)]]
   def __init__(self): self.tensor_cores = MetalRenderer.tensor_cores if hasattr(os, 'uname') and os.uname().machine == "arm64" else []
 
   # language options