diff --git a/.github/workflows/cubed.yml b/.github/workflows/cubed.yml
new file mode 100644
index 000000000..2b9cde4e8
--- /dev/null
+++ b/.github/workflows/cubed.yml
@@ -0,0 +1,33 @@
+name: Cubed
+
+on:
+  push:
+  pull_request:
+  # manual trigger
+  workflow_dispatch:
+
+jobs:
+  build:
+    # This workflow only runs on the origin org
+    # if: github.repository_owner == 'sgkit-dev'
+    runs-on: ubuntu-latest
+    strategy:
+      matrix:
+        python-version: ["3.11"]
+
+    steps:
+    - uses: actions/checkout@v4
+    - name: Set up Python ${{ matrix.python-version }}
+      uses: actions/setup-python@v5
+      with:
+        python-version: ${{ matrix.python-version }}
+
+    - name: Install deps and sgkit
+      run: |
+        python -m pip install --upgrade pip
+        python -m pip install -r requirements.txt -r requirements-dev.txt
+        python -m pip install -U git+https://github.com/cubed-dev/cubed.git -U git+https://github.com/cubed-dev/cubed-xarray.git -U git+https://github.com/pydata/xarray.git
+
+    - name: Test with pytest
+      run: |
+        pytest -v sgkit/tests/test_aggregation.py -k "test_count_call_alleles" --use-cubed
diff --git a/conftest.py b/conftest.py
index bcccb6e8a..9343393f3 100644
--- a/conftest.py
+++ b/conftest.py
@@ -2,6 +2,30 @@
 collect_ignore_glob = ["benchmarks/**", "sgkit/io/vcf/*.py", ".github/scripts/*.py"]
 
 
+def pytest_addoption(parser):
+    parser.addoption(
+        "--use-cubed", action="store_true", default=False, help="run with cubed"
+    )
+
+
+def use_cubed():
+    import dask
+    import xarray as xr
+
+    # set xarray to use cubed by default
+    xr.set_options(chunk_manager="cubed")
+
+    # ensure that dask compute raises if it is ever called
+    class AlwaysRaiseScheduler:
+        def __call__(self, dsk, keys, **kwargs):
+            raise RuntimeError("Dask 'compute' was called")
+
+    dask.config.set(scheduler=AlwaysRaiseScheduler())
+
+
 def pytest_configure(config) -> None:  # type: ignore
     # Add "gpu" marker
     config.addinivalue_line("markers", "gpu:Run tests that run on GPU")
+
+    if config.getoption("--use-cubed"):
+        use_cubed()
diff --git a/sgkit/stats/aggregation.py b/sgkit/stats/aggregation.py
index 41adcc221..8bbb738b5 100644
--- a/sgkit/stats/aggregation.py
+++ b/sgkit/stats/aggregation.py
@@ -76,17 +76,24 @@ def count_call_alleles(
 
     variables.validate(ds, {call_genotype: variables.call_genotype_spec})
     n_alleles = ds.sizes["alleles"]
-    G = da.asarray(ds[call_genotype])
-    shape = (G.chunks[0], G.chunks[1], n_alleles)
     # use numpy array to avoid dask task dependencies between chunks
     N = np.empty(n_alleles, dtype=np.uint8)
+    AC = xr.apply_ufunc(
+        count_alleles,
+        ds[call_genotype],
+        N,
+        input_core_dims=[["ploidy"], ["alleles"]],
+        output_core_dims=[["alleles"]],
+        exclude_dims={"ploidy"},
+        dask="parallelized",
+        dask_gufunc_kwargs=dict(allow_rechunk=False),
+        output_dtypes=np.uint8,
+    )
     new_ds = create_dataset(
         {
             variables.call_allele_count: (
                 ("variants", "samples", "alleles"),
-                da.map_blocks(
-                    count_alleles, G, N, chunks=shape, drop_axis=2, new_axis=2
-                ),
+                AC.data,
             )
         }
     )
diff --git a/sgkit/tests/test_aggregation.py b/sgkit/tests/test_aggregation.py
index a5e68645e..d4d3db1e5 100644
--- a/sgkit/tests/test_aggregation.py
+++ b/sgkit/tests/test_aggregation.py
@@ -139,8 +139,10 @@ def test_count_variant_alleles__chunked(using):
     calls = rs.randint(0, 1, size=(50, 10, 2))
     ds = get_dataset(calls)
     ac1 = count_variant_alleles(ds, using=using)
-    # Coerce from numpy to multiple chunks in all dimensions
-    ds["call_genotype"] = ds["call_genotype"].chunk(chunks=(5, 5, 1))
+    # Coerce from numpy to multiple chunks in all non-core dimensions
+    ds["call_genotype"] = ds["call_genotype"].chunk(
+        chunks={"variants": 5, "samples": 5}
+    )
     ac2 = count_variant_alleles(ds, using=using)
     assert isinstance(ac2["variant_allele_count"].data, da.Array)
     xr.testing.assert_equal(ac1, ac2)
@@ -265,10 +267,12 @@ def test_count_call_alleles__chunked():
     calls = rs.randint(0, 1, size=(50, 10, 2))
     ds = get_dataset(calls)
     ac1 = count_call_alleles(ds)
-    # Coerce from numpy to multiple chunks in all dimensions
-    ds["call_genotype"] = ds["call_genotype"].chunk(chunks=(5, 5, 1))
+    # Coerce from numpy to multiple chunks in all non-core dimensions
+    ds["call_genotype"] = ds["call_genotype"].chunk(
+        chunks={"variants": 5, "samples": 5}
+    )
     ac2 = count_call_alleles(ds)
-    assert isinstance(ac2["call_allele_count"].data, da.Array)
+    assert hasattr(ac2["call_allele_count"].data, "chunks")
     xr.testing.assert_equal(ac1, ac2)
 
 
diff --git a/sgkit/tests/test_ibs.py b/sgkit/tests/test_ibs.py
index e53fab7b7..c68a05569 100644
--- a/sgkit/tests/test_ibs.py
+++ b/sgkit/tests/test_ibs.py
@@ -1,6 +1,5 @@
 import pathlib
 
-import dask.array as da
 import numpy as np
 import pytest
 
@@ -42,13 +41,9 @@ def test_identity_by_state__diploid_biallelic(method, chunks, skipna):
         pass
     elif method == "frequencies":
         ds = count_call_alleles(ds)
-        ds["call_allele_count"] = (
-            ds.call_allele_count.dims,
-            ds.call_allele_count.data.rechunk(chunks),
-        )
+        ds["call_allele_count"] = ds["call_allele_count"].chunk(chunks)
     else:
-        gt = da.array(ds.call_genotype.data)
-        ds["call_genotype"] = ds.call_genotype.dims, gt.rechunk(chunks)
+        ds["call_genotype"] = ds["call_genotype"].chunk(chunks)
     ds = identity_by_state(ds, method=method, skipna=skipna)
     actual = ds.stat_identity_by_state.values
     expect = np.nanmean(
@@ -93,13 +88,9 @@ def test_identity_by_state__tetraploid_multiallelic(method, chunks, skipna):
     if chunks is None:
         pass
     elif method == "frequencies":
-        ds["call_allele_count"] = (
-            ds.call_allele_count.dims,
-            ds.call_allele_count.data.rechunk(chunks),
-        )
+        ds["call_allele_count"] = ds["call_allele_count"].chunk(chunks)
     else:
-        gt = da.array(ds.call_genotype.data)
-        ds["call_genotype"] = ds.call_genotype.dims, gt.rechunk(chunks)
+        ds["call_genotype"] = ds["call_genotype"].chunk(chunks)
     ds = identity_by_state(ds, method=method, skipna=skipna)
     actual = ds.stat_identity_by_state.values
     if skipna:
@@ -147,7 +138,7 @@ def test_identity_by_state__reference_implementation(ploidy, method, chunks, ski
         seed=0,
     )
     ds = call_allele_frequencies(ds)
-    ds.chunk(variants=chunks[0], samples=chunks[1], alleles=chunks[2])
+    ds = ds.chunk(variants=chunks[0], samples=chunks[1], alleles=chunks[2])
     # reference implementation
     AF = ds.call_allele_frequency.data
     if skipna:
diff --git a/sgkit/tests/test_popgen.py b/sgkit/tests/test_popgen.py
index 50fc9bb4c..32a9a8e6a 100644
--- a/sgkit/tests/test_popgen.py
+++ b/sgkit/tests/test_popgen.py
@@ -533,7 +533,7 @@ def test_Garud_h__raise_on_no_windows():
 
 
 @pytest.mark.filterwarnings("ignore::RuntimeWarning")
-@pytest.mark.parametrize("chunks", [((4,), (6,), (4,)), ((2, 2), (3, 3), (2, 2))])
+@pytest.mark.parametrize("chunks", [((4,), (6,), (4,)), ((2, 2), (3, 3), (4,))])
 def test_observed_heterozygosity(chunks):
     ds = simulate_genotype_call_dataset(
         n_variant=4,
@@ -599,7 +599,7 @@ def test_observed_heterozygosity(chunks):
 
 
 @pytest.mark.filterwarnings("ignore::RuntimeWarning")
-@pytest.mark.parametrize("chunks", [((4,), (6,), (4,)), ((2, 2), (3, 3), (2, 2))])
+@pytest.mark.parametrize("chunks", [((4,), (6,), (4,)), ((2, 2), (3, 3), (4,))])
 @pytest.mark.parametrize(
     "cohorts,expectation",
     [