diff --git a/.github/workflows/test.yml b/.github/workflows/test.yml
index a5a04c1929..258288bd58 100644
--- a/.github/workflows/test.yml
+++ b/.github/workflows/test.yml
@@ -43,7 +43,7 @@ jobs:
         PYTHONPATH=. DEBUG=2 EMULATE_AMD=1 FORWARD_ONLY=1 PYTHON=1 N=16 HALF=1 ACC_HALF=1 python3 ./extra/gemm/simple_matmul.py
         PYTHONPATH=. DEBUG=2 EMULATE_AMD=1 FORWARD_ONLY=1 PYTHON=1 N=64 HALF=1 ACC_HALF=1 python3 ./extra/gemm/simple_matmul.py
     - name: Test emulated CUDA tensor cores
-      run: DEBUG=2 EMULATE_CUDA=1 FORWARD_ONLY=1 PYTHON=1 python3 test/test_ops.py TestOps.test_gemm
+      run: DEBUG=2 EMULATE_CUDA=1 FORWARD_ONLY=1 PYTHON=1 python3 test/test_ops.py TestOps.test_gemm_fp16
     - name: Full test tensor cores
       run: |
         PYTHONPATH=. DEBUG=2 EMULATE_METAL=1 FORWARD_ONLY=1 PYTHON=1 python3 ./test/test_linearizer.py TestLinearizer.test_tensor_cores
diff --git a/test/test_ops.py b/test/test_ops.py
index 43e2d3795e..24bb3d6798 100644
--- a/test/test_ops.py
+++ b/test/test_ops.py
@@ -795,10 +795,13 @@ class TestOps(unittest.TestCase):
                                                                          np.arange(64,128,dtype=np.float32).reshape(8,8)])
   def test_small_gemm_eye(self):
     helper_test_op(None, lambda x,y: x.matmul(y), lambda x,y: x@y, vals=[np.eye(8).astype(np.float32), np.eye(8).astype(np.float32)])
+  @unittest.skipIf(Device.DEFAULT in ["NV", "LLVM", "GPU", "CUDA"], "not supported on these in CI")
+  def test_gemm_fp16(self):
+    helper_test_op([(64,64), (64,64)], lambda x,y: x.half().matmul(y.half()))
   def test_gemm(self):
-    helper_test_op([(64,64), (64,64)], lambda x,y: x.matmul(y), Tensor.dot)
+    helper_test_op([(64,64), (64,64)], lambda x,y: x.matmul(y))
   def test_big_gemm(self):
-    helper_test_op([(256,256), (256,256)], lambda x,y: x.matmul(y), Tensor.dot, atol=1e-4)
+    helper_test_op([(256,256), (256,256)], lambda x,y: x.matmul(y), atol=1e-4)
   @unittest.skipIf(IMAGE>0, "no 0 in shape matmul on images")
   def test_gemm_with_zeros_shape(self):
     helper_test_op([(8,8), (8,0)], lambda x,y: x.matmul(y), Tensor.dot, atol=1e-7)