diff --git a/.github/workflows/codspeed.yml b/.github/workflows/codspeed.yml
index fed68f03..4f0aee45 100644
--- a/.github/workflows/codspeed.yml
+++ b/.github/workflows/codspeed.yml
@@ -25,4 +25,9 @@ jobs:
       - name: Run benchmarks
         uses: CodSpeedHQ/action@v3
         with:
-          run: pytest benchmarks/ --codspeed
+          run: SPARSE_BACKEND=Numba pytest benchmarks/ --codspeed
+
+      - name: Run benchmarks
+        uses: CodSpeedHQ/action@v3
+        with:
+          run: SPARSE_BACKEND=Finch pytest benchmarks/ --codspeed
diff --git a/benchmarks/conftest.py b/benchmarks/conftest.py
index b80d9b88..bf5e12a8 100644
--- a/benchmarks/conftest.py
+++ b/benchmarks/conftest.py
@@ -1,3 +1,5 @@
+import sparse
+
 import pytest
 
 
@@ -6,6 +8,21 @@ def seed(scope="session"):
     return 42
 
 
+def get_backend_id(param):
+    backend = param
+    return f"{backend=}"
+
+
+@pytest.fixture(params=[sparse._BACKEND.value], autouse=True, ids=get_backend_id)
+def backend(request):
+    return request.param
+
+
+@pytest.fixture
+def min_size(scope="session"):
+    return 50
+
+
 @pytest.fixture
 def max_size(scope="session"):
     return 2**26
diff --git a/benchmarks/test_benchmark_coo.py b/benchmarks/test_benchmark_coo.py
index 108ac075..7bc33a63 100644
--- a/benchmarks/test_benchmark_coo.py
+++ b/benchmarks/test_benchmark_coo.py
@@ -15,16 +15,19 @@ def format_id(format):
 
 
 @pytest.mark.parametrize("format", ["coo", "gcxs"])
-def test_matmul(benchmark, sides, format, seed, max_size, ids=format_id):
+def test_matmul(benchmark, sides, seed, format, backend, min_size, max_size, ids=format_id):
     m, n, p = sides
 
-    if m * n >= max_size or n * p >= max_size:
+    if m * n >= max_size or n * p >= max_size or m * n <= min_size or n * p <= min_size:
         pytest.skip()
 
     rng = np.random.default_rng(seed=seed)
     x = sparse.random((m, n), density=DENSITY, format=format, random_state=rng)
     y = sparse.random((n, p), density=DENSITY, format=format, random_state=rng)
 
+    if hasattr(sparse, "compiled"):
+        operator.matmul = sparse.compiled(operator.matmul)
+
     x @ y  # Numba compilation
 
     @benchmark
@@ -50,8 +53,12 @@ def elemwise_args(request, seed, max_size):
 
 
 @pytest.mark.parametrize("f", [operator.add, operator.mul])
-def test_elemwise(benchmark, f, elemwise_args):
+def test_elemwise(benchmark, f, elemwise_args, backend):
     x, y = elemwise_args
+
+    if hasattr(sparse, "compiled"):
+        f = sparse.compiled(f)
+
     f(x, y)
 
     @benchmark
@@ -78,6 +85,10 @@ def elemwise_broadcast_args(request, seed, max_size):
 @pytest.mark.parametrize("f", [operator.add, operator.mul])
 def test_elemwise_broadcast(benchmark, f, elemwise_broadcast_args):
     x, y = elemwise_broadcast_args
+
+    if hasattr(sparse, "compiled"):
+        f = sparse.compiled(f)
+
     f(x, y)
 
     @benchmark
@@ -101,6 +112,9 @@ def test_index_scalar(benchmark, indexing_args):
     side = x.shape[0]
     rank = x.ndim
 
+    if hasattr(sparse, "compiled"):
+        operator.getitem = sparse.compiled(operator.getitem)
+
     x[(side // 2,) * rank]  # Numba compilation
 
     @benchmark
@@ -113,6 +127,9 @@ def test_index_slice(benchmark, indexing_args):
     side = x.shape[0]
     rank = x.ndim
 
+    if hasattr(sparse, "compiled"):
+        operator.getitem = sparse.compiled(operator.getitem)
+
     x[(slice(side // 2),) * rank]  # Numba compilation
 
     @benchmark
@@ -126,6 +143,9 @@ def test_index_fancy(benchmark, indexing_args, seed):
     rng = np.random.default_rng(seed=seed)
     index = rng.integers(0, side, size=(side // 2,))
 
+    if hasattr(sparse, "compiled"):
+        operator.getitem = sparse.compiled(operator.getitem)
+
     x[index]  # Numba compilation
 
     @benchmark
@@ -165,6 +185,9 @@ def densemul_args(request, sides, seed, max_size):
 def test_gcxs_dot_ndarray(benchmark, densemul_args):
     x, t = densemul_args
 
+    if hasattr(sparse, "compiled"):
+        operator.matmul = sparse.compiled(operator.matmul)
+
     # Numba compilation
     x @ t