diff --git a/.github/workflows/benchmark.yml b/.github/workflows/benchmark.yml
new file mode 100644
index 0000000..c735fd4
--- /dev/null
+++ b/.github/workflows/benchmark.yml
@@ -0,0 +1,60 @@
+name: Benchmark
+
+on:
+  push:
+    branches: [main]
+  pull_request: {}
+  workflow_dispatch: {}
+
+permissions:
+  contents: read
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.job }}-${{ github.ref }}
+  cancel-in-progress: true
+
+defaults:
+  run:
+    shell: bash -e -l {0}
+
+jobs:
+  benchmarks:
+    name: Linux - ${{ matrix.sys.name }}
+    runs-on: namespace-profile-benchmark;container.privileged=true;container.host-pid-namespace=true
+    strategy:
+      matrix:
+        sys:
+          - { name: "sse4.2", flags: "-march=x86-64-v2" }
+          - { name: "avx2",   flags: "-march=x86-64-v3" }
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v6
+      - name: Set conda environment
+        uses: mamba-org/setup-micromamba@v2
+        with:
+          environment-name: myenv
+          create-args: >-
+            cmake
+            ninja
+            benchmark
+            xsimd=14.3.0
+          init-shell: bash
+          cache-downloads: true
+      - name: Build benchmarks
+        run: |
+          cmake -G Ninja -B build/ \
+            -D CMAKE_BUILD_TYPE="RelWithDebInfo" \
+            -D CMAKE_CXX_FLAGS="${{ matrix.sys.flags }}" \
+            -D BUILD_BENCHMARK=ON
+          cmake --build build --target benchmark_xsimd_algorithm --parallel 8
+      - name: Run benchmarks
+        run: |
+          ./build/benchmark/benchmark_xsimd_algorithm \
+            --benchmark_repetitions=3 \
+            --benchmark_out=benchmark-${{ matrix.sys.name }}.json \
+            --benchmark_out_format=json
+      - name: Upload results
+        uses: actions/upload-artifact@v4
+        with:
+          name: benchmark-${{ github.event.pull_request.head.sha || github.sha }}-${{ matrix.sys.name }}
+          path: benchmark-${{ matrix.sys.name }}.json
diff --git a/benchmark/CMakeLists.txt b/benchmark/CMakeLists.txt
index e214e36..ad27859 100644
--- a/benchmark/CMakeLists.txt
+++ b/benchmark/CMakeLists.txt
@@ -16,12 +16,25 @@ endif ()
 
 find_package(benchmark REQUIRED)
 
-set(XSIMD_ALGORITHM_BENCHMARKS
+# Regular benchmark of algorithm implemented here
+add_executable(
+    benchmark_xsimd_algorithm
     main.cpp
-    bench_math.cpp
     bench_map.cpp
 )
+target_link_libraries(
+    benchmark_xsimd_algorithm
+    PRIVATE xsimd-algorithm xsimd::test-utils benchmark::benchmark
+)
 
-add_executable(benchmark_xsimd_algorithm ${XSIMD_ALGORITHM_BENCHMARKS})
-target_link_libraries(benchmark_xsimd_algorithm
-    PRIVATE xsimd-algorithm xsimd::test-utils benchmark::benchmark)
+# Benchmark actually targeting xsimd functionalities (hot loop in L1 cache,
+# no header and trailers). Should probably be moved to xsimd itself.
+add_executable(
+    benchmark_xsimd
+    main.cpp
+    bench_math.cpp
+)
+target_link_libraries(
+    benchmark_xsimd
+    PRIVATE xsimd-algorithm xsimd::test-utils benchmark::benchmark
+)
diff --git a/benchmark/bench_map.cpp b/benchmark/bench_map.cpp
index 562c53a..44cea96 100644
--- a/benchmark/bench_map.cpp
+++ b/benchmark/bench_map.cpp
@@ -145,13 +145,9 @@ namespace
     bool const registered = []
     {
         register_benches<xsimd::test::sqrt_op<float>>();
-        register_benches<xsimd::test::sqrt_op<double>>();
-        register_benches<xsimd::test::abs_op<float>>();
         register_benches<xsimd::test::abs_op<double>>();
         register_benches<xsimd::test::exp_op<float>>();
-        register_benches<xsimd::test::exp_op<double>>();
         register_binary_benches<xsimd::test::add_op<std::int32_t>>();
-        register_binary_benches<xsimd::test::multiply_op<float>>();
         register_binary_benches<xsimd::test::mixed_multiply_op>();
         return true;
     }();
diff --git a/benchmark/bench_utils.hpp b/benchmark/bench_utils.hpp
index f7b3686..586eb41 100644
--- a/benchmark/bench_utils.hpp
+++ b/benchmark/bench_utils.hpp
@@ -22,14 +22,9 @@ namespace xsimd::bench
     {
         constexpr auto batch_size = static_cast<std::int64_t>(xsimd::batch<T>::size);
 
-        auto sizes = std::vector<std::int64_t> {};
-        for (std::int64_t size : { 64, 1024, 65536, 1 << 21 })
-        {
-            auto const whole = size - (size % batch_size);
-            sizes.push_back(whole);
-            sizes.push_back(whole + batch_size / 2 + 1);
-        }
-        return sizes;
+        return std::vector<std::int64_t> {
+            64, 67, 1024, 1027, 65536, 2097152
+        };
     }
 
     template <typename T>