diff --git a/.gitlab-ci.yml b/.gitlab-ci.yml
index a3d41d36437..b61f87d8f64 100644
--- a/.gitlab-ci.yml
+++ b/.gitlab-ci.yml
@@ -26,6 +26,8 @@ include:
   - local: '.gitlab/add-interrupt.yml'
     rules:
       - if: $CI_COMMIT_BRANCH != "master" && $CI_COMMIT_BRANCH != "develop" && $CI_COMMIT_TAG !~ /^v\d+\.\d+\.\d+/
+  # build.yml use the above stuff
+  - local: '.gitlab/build.yml'
 
 sync:
   stage: sync
@@ -91,541 +93,6 @@ trigger_pipeline:
       fi
 
 
-# cuda 11.4 and friends
-build/cuda110/nompi/gcc/cuda/release/shared:
-  extends:
-    - .build_and_test_template
-    - .default_variables
-    - .quick_test_condition
-    - .use_gko_cuda110-mvapich-gnu9-llvm9
-  variables:
-    BUILD_OMP: "ON"
-    BUILD_CUDA: "ON"
-    BUILD_MPI: "OFF"
-    BUILD_TYPE: "Release"
-    FAST_TESTS: "ON"
-    # fix gtest issue https://github.com/google/googletest/issues/3514
-    CXX_FLAGS: "-Wno-error=maybe-uninitialized"
-    # disable spurious unused argument warning
-    # this is seemingly broken with CUDA 11
-    # EXTRA_CMAKE_FLAGS: "-DCMAKE_CUDA_FLAGS=-diag-suppress=177"
-
-build/nvhpc227/cuda117/nompi/nvcpp/debug/shared:
-  extends:
-    - .build_and_test_template
-    - .default_variables
-    - .full_test_condition
-    - .use_gko_nvhpc227-cuda117-openmpi-gnu11-llvm14
-  variables:
-    CXX_COMPILER: "nvc++"
-    BUILD_OMP: "ON"
-    BUILD_CUDA: "ON"
-    BUILD_MPI: "OFF"
-    BUILD_TYPE: "Debug"
-    FAST_TESTS: "ON"
-    # disable common gflags warnings
-    CXX_FLAGS: "--diag_suppress=useless_using_declaration,declared_but_not_referenced"
-    # disable spurious unused argument warning
-    EXTRA_CMAKE_FLAGS: "-DCMAKE_CUDA_FLAGS=-diag-suppress=177"
-
-build/cuda118/nompi/gcc/cuda/debug/shared:
-  extends:
-    - .build_and_test_tum_template
-    - .default_variables
-    - .full_test_condition
-    - .use_tum-nvidia
-  variables:
-    BUILD_CUDA: "ON"
-    BUILD_HWLOC: "OFF"
-    ENABLE_HALF: "ON"
-    BUILD_TYPE: "Debug"
-    MODULE_LOAD: "cmake/3.18.6 cuda/11.8.0 gcc/7.5.0"
-
-build/cuda120/openmpi/gcc/cuda/release/static:
-  extends:
-    - .build_and_test_tum_template
-    - .default_variables
-    - .full_test_condition
-    - .use_tum-nvidia
-  variables:
-    BUILD_CUDA: "ON"
-    BUILD_HWLOC: "OFF"
-    ENABLE_HALF: "ON"
-    BUILD_MPI: "ON"
-    BUILD_SHARED_LIBS: "OFF"
-    BUILD_TYPE: "Release"
-    MODULE_LOAD: "cmake/3.18.6 cuda/12.0.1 gcc/12.4.0 openmpi/4.1.8"
-
-build/cuda122/openmpi/gcc/cuda/release/shared:
-  extends:
-    - .build_and_test_tum_template
-    - .default_variables
-    - .full_test_condition
-    - .use_tum-nvidia
-  variables:
-    BUILD_CUDA: "ON"
-    BUILD_HWLOC: "OFF"
-    ENABLE_HALF: "ON"
-    BUILD_TYPE: "Release"
-    MODULE_LOAD: "cmake/3.20.6 cuda/12.2.2 gcc/12.4.0 openmpi/5.0.7"
-
-build/cuda124/mpich/gcc/cuda/release/shared:
-  extends:
-    - .build_and_test_tum_template
-    - .default_variables
-    - .full_test_condition
-    - .use_tum-nvidia
-  variables:
-    BUILD_CUDA: "ON"
-    BUILD_HWLOC: "OFF"
-    BUILD_MPI: "ON"
-    ENABLE_HALF: "ON"
-    ENABLE_BFLOAT16: "ON"
-    BUILD_TYPE: "Release"
-    MODULE_LOAD: "cmake/3.30.8 cuda/12.6.3 gcc/13.3.0 mpich"
-
-build/cuda126/nompi/gcc/cuda/release/shared:
-  extends:
-    - .build_and_test_tum_template
-    - .default_variables
-    - .full_test_condition
-    - .use_tum-nvidia
-  variables:
-    BUILD_CUDA: "ON"
-    BUILD_HWLOC: "OFF"
-    ENABLE_HALF: "ON"
-    ENABLE_BFLOAT16: "ON"
-    BUILD_TYPE: "Release"
-    MODULE_LOAD: "cmake/3.30.8 cuda/12.6.3 gcc/13.3.0"
-
-build/cuda130/nompi/gcc/cuda/release/shared:
-  extends:
-    - .build_and_test_tum_template
-    - .default_variables
-    - .full_test_condition
-    - .use_tum-nvidia
-  variables:
-    BUILD_CUDA: "ON"
-    BUILD_HWLOC: "OFF"
-    ENABLE_HALF: "ON"
-    ENABLE_BFLOAT16: "ON"
-    BUILD_TYPE: "Release"
-    MODULE_LOAD: "cmake/3.30.8 cuda/13.0.2 gcc/14.3.0"
-
-build/cuda131/nompi/gcc/cuda/release/shared:
-  extends:
-    - .build_and_test_tum_template
-    - .default_variables
-    - .full_test_condition
-    - .use_tum-nvidia
-  variables:
-    BUILD_CUDA: "ON"
-    BUILD_HWLOC: "OFF"
-    ENABLE_HALF: "ON"
-    ENABLE_BFLOAT16: "ON"
-    BUILD_TYPE: "Release"
-    MODULE_LOAD: "cmake/3.30.8 cuda/13.1.1 gcc/14.3.0"
-
-build/amd/openmpi/gcc/rocm620/debug/shared:
-  extends:
-    - .build_and_test_tum_template
-    - .default_variables
-    - .quick_test_condition
-    - .use_tum-amd
-  variables:
-    BUILD_HIP: "ON"
-    BUILD_HWLOC: "OFF"
-    BUILD_MPI: "ON"
-    BUILD_OMP: "ON"
-    BUILD_TYPE: "Debug"
-    BUILD_SHARED_LIBS: "ON"
-    MODULE_LOAD: "cmake/3.24.4 rocm/6.2.0 gcc/13.3.0 openmpi/5.0.7"
-
-# rocm 6.3.x is the last version support MI50 officially
-build/amd/openmpi/gcc/rocm634_wo_omp/release/shared:
-  extends:
-    - .build_and_test_tum_template
-    - .default_variables
-    - .full_test_condition
-    - .use_tum-amd
-  variables:
-    BUILD_HIP: "ON"
-    BUILD_HWLOC: "OFF"
-    BUILD_MPI: "ON"
-    BUILD_TYPE: "Release"
-    MODULE_LOAD: "cmake/3.29.6 rocm/6.3.4 gcc/13.3.0 openmpi/5.0.7"
-
-# mi50 is not officially supported by ROCm >= 6.4.0
-build/amd/nompi/gcc/rocm644/debug/static:
-  extends:
-    - .build_and_test_tum_template
-    - .default_variables
-    - .full_test_condition
-    - .use_tum-amd-mi210
-  variables:
-    BUILD_HIP: "ON"
-    BUILD_HWLOC: "OFF"
-    BUILD_MPI: "OFF"
-    BUILD_OMP: "OFF"
-    BUILD_TYPE: "Debug"
-    BUILD_SHARED_LIBS: "OFF"
-    MODULE_LOAD: "cmake/3.29.6 rocm/6.4.4 gcc/14.3.0"
-
-build/amd/nompi/gcc/rocm710/release/shared:
-  extends:
-    - .build_and_test_tum_template
-    - .default_variables
-    - .full_test_condition
-    - .use_tum-amd-mi210
-  variables:
-    BUILD_HIP: "ON"
-    BUILD_HWLOC: "OFF"
-    BUILD_MPI: "OFF"
-    BUILD_OMP: "OFF"
-    BUILD_TYPE: "Release"
-    MODULE_LOAD: "cmake/3.29.6 rocm/7.1.0 gcc/14.3.0"
-
-build/amd/nompi/gcc/rocm720/release/shared:
-  extends:
-    - .build_and_test_tum_template
-    - .default_variables
-    - .full_test_condition
-    - .use_tum-amd
-  variables:
-    BUILD_HIP: "ON"
-    BUILD_OMP: "OFF"
-    BUILD_HWLOC: "OFF"
-    BUILD_TYPE: "Release"
-    ENABLE_BFLOAT16: "OFF"
-    ENABLE_HALF: "OFF"
-    MODULE_LOAD: "cmake/3.25.3 rocm/7.2.0 gcc/12.4.0"
-
-# no cuda but latest gcc and clang
-build/nocuda/nompi/gcc/core/debug/static:
-  extends:
-    - .build_and_test_template
-    - .default_variables
-    - .full_test_condition
-    - .use_gko-nocuda-nompi-gnu9-llvm8
-  variables:
-    BUILD_TYPE: "Debug"
-    FAST_TESTS: "ON"
-    BUILD_SHARED_LIBS: "OFF"
-    BUILD_HWLOC: "OFF"
-
-build/nocuda/nompi/clang/core/release/shared:
-  extends:
-    - .build_and_test_template
-    - .default_variables
-    - .full_test_condition
-    - .use_gko-nocuda-nompi-gnu9-llvm8
-  variables:
-    CXX_COMPILER: "clang++"
-    BUILD_TYPE: "Release"
-    ENABLE_HALF: "OFF"
-    ENABLE_BFLOAT16: "ON"
-
-build/nocuda/nompi/gcc/omp/release/shared:
-  extends:
-    - .build_and_test_template
-    - .default_variables
-    - .quick_test_condition
-    - .use_gko-nocuda-nompi-gnu9-llvm8
-  variables:
-    BUILD_OMP: "ON"
-    BUILD_TYPE: "Release"
-    ENABLE_HALF: "ON"
-    ENABLE_BFLOAT16: "ON"
-
-build/nocuda/openmpi/clang/omp/debug/static:
-  extends:
-    - .build_and_test_template
-    - .default_variables
-    - .full_test_condition
-    - .use_gko-nocuda-openmpi-gnu9-llvm8
-  variables:
-    CXX_COMPILER: "clang++"
-    BUILD_OMP: "ON"
-    MPI_AS_ROOT: "ON"
-    BUILD_MPI: "ON"
-    BUILD_TYPE: "Debug"
-    FAST_TESTS: "ON"
-    BUILD_SHARED_LIBS: "OFF"
-
-build/nocuda/openmpi/clang/omp/glibcxx-debug-release/shared:
-  extends:
-    - .build_and_test_template
-    - .default_variables
-    - .full_test_condition
-    - .use_gko-nocuda-nompi-gnu9-llvm8
-  variables:
-    CXX_COMPILER: "clang++"
-    BUILD_OMP: "ON"
-    MPI_AS_ROOT: "ON"
-    BUILD_MPI: "ON"
-    CXX_FLAGS: "-Wpedantic -D_GLIBCXX_DEBUG=1"
-    # The tests are prohibitively slow in Debug
-    BUILD_TYPE: "Release"
-
-# nocuda with old compiler
-build/nocuda/nompi/gcc/omp/release/static:
-  extends:
-    - .build_and_test_template
-    - .default_variables
-    - .full_test_condition
-    - .use_gko-nocuda-nompi-gnu9-llvm8
-  variables:
-    BUILD_OMP: "ON"
-    BUILD_TYPE: "Release"
-    BUILD_SHARED_LIBS: "OFF"
-    ENABLE_HALF: "OFF"
-    ENABLE_BFLOAT16: "ON"
-
-build/nocuda-nomixed/nompi/clang/omp/release/static:
-  extends:
-    - .build_and_test_template
-    - .default_variables
-    - .full_test_condition
-    - .use_gko-nocuda-nompi-gnu9-llvm8
-  variables:
-    CXX_COMPILER: "clang++"
-    BUILD_OMP: "ON"
-    BUILD_TYPE: "Release"
-    BUILD_SHARED_LIBS: "OFF"
-    MIXED_PRECISION: "OFF"
-
-build/nocuda-nomixed/openmpi/gcc/omp/release/shared:
-  extends:
-    - .build_and_test_template
-    - .default_variables
-    - .full_test_condition
-    - .use_gko-nocuda-nompi-gnu9-llvm8
-  variables:
-    BUILD_MPI: "ON"
-    MPI_AS_ROOT: "ON"
-    BUILD_OMP: "ON"
-    BUILD_TYPE: "Release"
-    MIXED_PRECISION: "OFF"
-
-build/nocuda-nomixed/nompi/clang/omp/debug/static:
-  extends:
-    - .build_and_test_template
-    - .default_variables
-    - .full_test_condition
-    - .use_gko-nocuda-nompi-gnu9-llvm8
-  variables:
-    CXX_COMPILER: "clang++"
-    BUILD_OMP: "ON"
-    BUILD_TYPE: "Debug"
-    BUILD_SHARED_LIBS: "OFF"
-    MIXED_PRECISION: "OFF"
-
-build/nogpu/nompi/gcc/omp/debug/static:
-  extends:
-    - .build_and_test_tum_template
-    - .default_variables
-    - .full_test_condition
-    - .use_tum-cpu
-  variables:
-    BUILD_OMP: "ON"
-    BUILD_TYPE: "Debug"
-    BUILD_SHARED_LIBS: "OFF"
-    MIXED_PRECISION: "OFF"
-    BUILD_HWLOC: "OFF"
-    MODULE_LOAD: "cmake/3.18.6 gcc/7.5.0"
-
-build/nogpu/nompi/gcc/omp/release/static:
-  extends:
-    - .build_and_test_tum_template
-    - .default_variables
-    - .full_test_condition
-    - .use_tum-cpu
-  variables:
-    BUILD_OMP: "ON"
-    BUILD_TYPE: "Release"
-    BUILD_SHARED_LIBS: "OFF"
-    ENABLE_HALF: "ON"
-    ENABLE_BFLOAT16: "ON"
-    BUILD_HWLOC: "OFF"
-    MODULE_LOAD: "cmake/3.18.6 gcc/9.5.0"
-
-build/nogpu/nompi/gcc/noomp/release/shared:
-  extends:
-    - .build_and_test_tum_template
-    - .default_variables
-    - .full_test_condition
-    - .use_tum-cpu
-  variables:
-    BUILD_TYPE: "Release"
-    BUILD_HWLOC: "OFF"
-    MODULE_LOAD: "cmake/3.18.6 gcc/9.5.0"
-
-build/nogpu/nompi/gcc12/noomp/release/shared:
-  extends:
-    - .build_and_test_tum_template
-    - .default_variables
-    - .full_test_condition
-    - .use_tum-cpu
-  variables:
-    BUILD_TYPE: "Release"
-    BUILD_HWLOC: "OFF"
-    MODULE_LOAD: "cmake/3.18.6 gcc/12.4.0"
-
-build/nogpu/mpi/gcc13/noomp/release/shared:
-  extends:
-    - .build_and_test_tum_template
-    - .default_variables
-    - .quick_test_condition
-    - .use_tum-cpu
-  variables:
-    BUILD_TYPE: "Release"
-    BUILD_HWLOC: "OFF"
-    BUILD_MPI: "ON"
-    MODULE_LOAD: "cmake/3.18.6 gcc/13.3.0 openmpi"
-
-# spack oneapi 2023.1
-build/icpx20231/gpu/release/shared:
-  extends:
-    - .build_and_test_template
-    - .default_variables
-    - .quick_test_condition
-    - .use_gko-oneapi20231-gpu
-  variables:
-    CXX_COMPILER: "icpx"
-    CXX_FLAGS: "-Wpedantic -ffp-model=precise"
-    BUILD_SYCL: "ON"
-    BUILD_TYPE: "Release"
-    BUILD_SHARED_LIBS: "ON"
-    DPCPP_SINGLE_MODE: "ON"
-    ONEAPI_DEVICE_SELECTOR: "opencl:gpu"
-    BUILD_HWLOC: "OFF"
-
-build/dpcpp20231/gpu/release/shared:
-  extends:
-    - .build_and_test_tum_template
-    - .default_variables
-    - .full_test_condition
-    - .use_tum-intel-fp32
-  variables:
-    CXX_COMPILER: "dpcpp"
-    CXX_FLAGS: "-Wpedantic -ffp-model=precise"
-    BUILD_SYCL: "ON"
-    BUILD_TYPE: "Release"
-    BUILD_SHARED_LIBS: "ON"
-    DPCPP_SINGLE_MODE: "ON"
-    MODULE_LOAD: "cmake/3.21.7 intel-oneapi-compilers/2023.1.0 intel-oneapi-dpl/2022.1.0 intel-oneapi-tbb/2021.9.0 intel-oneapi-mkl/2023.1.0"
-    ONEAPI_DEVICE_SELECTOR: "opencl:gpu"
-    BUILD_HWLOC: "OFF"
-
-build/icpx202402/gpu/release/shared:
-  extends:
-    - .build_and_test_tum_template
-    - .default_variables
-    - .full_test_condition
-    - .use_tum-intel
-  variables:
-    CXX_COMPILER: "icpx"
-    CXX_FLAGS: "-Wpedantic -ffp-model=precise -Wno-deprecated-declarations"
-    BUILD_SYCL: "ON"
-    BUILD_HWLOC: "OFF"
-    BUILD_TYPE: "Release"
-    ENABLE_HALF: "OFF"
-    ENABLE_BFLOAT16: "OFF"
-    ONEAPI_DEVICE_SELECTOR: "opencl:gpu"
-    MODULE_LOAD: "cmake/3.21.7 intel-oneapi-compilers/2024.0.2 intel-oneapi-dpl/2022.3.0 intel-oneapi-tbb/2021.11.0 intel-oneapi-mkl/2024.0.0"
-
-build/icpx202421/gpu/release/shared:
-  extends:
-    - .build_and_test_tum_template
-    - .default_variables
-    - .full_test_condition
-    - .use_tum-intel
-  variables:
-    CXX_COMPILER: "icpx"
-    CXX_FLAGS: "-Wpedantic -ffp-model=precise -Wno-deprecated-declarations"
-    BUILD_SYCL: "ON"
-    BUILD_OMP: "ON"
-    BUILD_HWLOC: "OFF"
-    BUILD_TYPE: "Release"
-    ENABLE_HALF: "ON"
-    ENABLE_BFLOAT16: "ON"
-    ONEAPI_DEVICE_SELECTOR: "level_zero:gpu"
-    MODULE_LOAD: "cmake/3.28.6 intel-oneapi-compilers/2024.2.1 intel-oneapi-dpl/2022.6.1 intel-oneapi-tbb/2021.13.1 intel-oneapi-mkl/2024.2.1"
-
-# TODO: Enable when debug shared library size issues are fixed
-# build/dpcpp/level_zero_igpu/debug/shared:
-#   extends:
-#     - .build_and_test_template
-#     - .default_variables
-#     - .full_test_condition
-#     - .use_gko-oneapi-gpu
-#   variables:
-#     CXX_COMPILER: "dpcpp"
-#     CXX_FLAGS: "-Wpedantic -ffp-model=precise"
-#     BUILD_SYCL: "ON"
-#     BUILD_TYPE: "Debug"
-#     BUILD_SHARED_LIBS: "ON"
-#     DPCPP_SINGLE_MODE: "ON"
-#     ONEAPI_DEVICE_SELECTOR: "level_zero:gpu"
-
-# It gives two available backends of GPU on tests
-build/dpcpp/gpu/release/shared:
-  extends:
-    - .build_and_test_template
-    - .default_variables
-    - .full_test_condition
-    - .use_gko-oneapi-gpu
-  variables:
-    CXX_COMPILER: "dpcpp"
-    CXX_FLAGS: "-Wpedantic -ffp-model=precise"
-    BUILD_SYCL: "ON"
-    BUILD_TYPE: "Release"
-    # static builds take too long
-    BUILD_SHARED_LIBS: "ON"
-    DPCPP_SINGLE_MODE: "ON"
-    ONEAPI_DEVICE_SELECTOR: "*:gpu"
-    BUILD_HWLOC: "OFF"
-
-# windows jobs
-# Note that this is using Powershell, not bash
-# Disable it now until the server is back
-# build/windows-cuda/release/shared:
-#   extends:
-#     - .quick_test_condition
-#   stage: build
-#   script:
-#     - if (Test-Path build) { rm -r -fo build }
-#     - if (Test-Path install) { rm -r -fo install }
-#     - mkdir build
-#     - mkdir install
-#     - nvidia-smi
-#     - cmake -B build -DBUILD_SHARED_LIBS=ON -DGINKGO_BUILD_CUDA=ON "-DCMAKE_INSTALL_PREFIX=$pwd\install" .
-#     - cmake --build build --config Release -j16
-#     - ctest --test-dir build -C Release --no-tests=error --output-on-failure
-#     - $env:PATH+=";$pwd/install/bin"
-#     - cmake --install build --config Release
-#     - cmake --build build --target test_install --config Release
-#   tags:
-#     - windows-cuda
-
-# build/windows/release/shared:
-#   extends:
-#     - .full_test_condition
-#   stage: build
-#   script:
-#     - if (Test-Path build) { rm -r -fo build }
-#     - if (Test-Path install) { rm -r -fo install }
-#     - mkdir build
-#     - mkdir install
-#     - cmake -B build -DBUILD_SHARED_LIBS=ON -DGINKGO_BUILD_CUDA=OFF "-DCMAKE_INSTALL_PREFIX=$pwd\install" .
-#     - cmake --build build --config Release -j16
-#     - ctest --test-dir build -C Release --no-tests=error --output-on-failure -j16
-#     - $env:PATH+=";$pwd/install/bin"
-#     - cmake --install build --config Release
-#     - cmake --build build --target test_install --config Release
-#   tags:
-#     - windows
-
 # Job with important warnings as error
 warnings:
   stage: code_quality
@@ -797,22 +264,6 @@ gh-pages:
   needs: []
 
 
-threadsanitizer:
-  stage: QoS_tools
-  extends:
-    - .default_variables
-    - .deploy_condition
-    - .before_script_template
-    - .use_gko-rocm514-nompi-gnu11-llvm11
-    - .disable_job_condition
-  script:
-    - LD_PRELOAD=/usr/local/lib/libomp.so
-      CC=clang CXX=clang++
-        ctest -V -S cmake/CTestScript.cmake -DCTEST_BUILD_CONFIGURATION=TSAN
-        -DCTEST_MEMORYCHECK_TYPE=ThreadSanitizer
-        -DCTEST_MEMORYCHECK_SANITIZER_OPTIONS=ignore_noninstrumented_modules=1
-        --timeout 6000
-
 leaksanitizer:
   stage: QoS_tools
   extends:
@@ -872,48 +323,3 @@ new-issue-on-failure:
       - develop
       - master
   dependencies: []
-
-
-## Benchmark SpMV
-benchmark-cuda-spmv-build:
-  extends:
-    - .build_template
-    - .default_variables
-    - .use_gko_cuda114-openmpi-gnu10-llvm12
-    - .benchmark-spmv-cuda-rules
-    - .disable_job_condition
-  stage: benchmark-build
-  variables:
-    BUILD_OMP: "ON"
-    BUILD_CUDA: "ON"
-    BUILD_MPI: "ON"
-    MPI_AS_ROOT: "ON"
-    BUILD_HIP: "OFF"
-    BUILD_TYPE: "Release"
-    CUDA_ARCH: 80
-    USE_NAME: "benchmark-cuda-spmv-${CI_PIPELINE_ID}"
-    KEEP_CONTAINER: "ON"
-    USE_SLURM: 0
-
-
-# The name must be of the same as above without `-build`
-benchmark-cuda-spmv:
-  extends:
-    - .benchmark_template
-    - .default_variables
-    - .use_gko_cuda114-openmpi-gnu10-llvm12
-    - .benchmark-spmv-cuda-rules
-    - .disable_job_condition
-  stage: benchmark-cuda
-  variables:
-    BENCHMARK_REPO: git@github.com:ginkgo-project/ginkgo-data.git
-    USE_NAME: "benchmark-cuda-spmv-${CI_PIPELINE_ID}"
-    SLURM_PARTITION: "accelerated"
-    SLURM_GRES: "gpu:1"
-    SLURM_EXCLUSIVE: "ON"
-    SLURM_TIME: "4:00:00"
-    BENCHMARK: "spmv"
-    EXECUTOR: "cuda"
-    SYSTEM_NAME: "A100"
-  dependencies: null
-  needs: [ "benchmark-cuda-spmv-build" ]
diff --git a/.gitlab/build.yml b/.gitlab/build.yml
new file mode 100644
index 00000000000..03bb536fe8b
--- /dev/null
+++ b/.gitlab/build.yml
@@ -0,0 +1,469 @@
+build/cuda110/nompi/gcc/cuda/release/shared:
+  extends:
+    - .build_and_test_template
+    - .default_variables
+    - .quick_test_condition
+    - .use_gko_cuda110-mvapich-gnu9-llvm9
+  variables:
+    BUILD_OMP: 'ON'
+    BUILD_CUDA: 'ON'
+    BUILD_MPI: 'OFF'
+    BUILD_TYPE: Release
+    FAST_TESTS: 'ON'
+    _COMMENTS: disabling spurious unused argument warning is seemingly broken with CUDA 11
+
+build/nvhpc227/cuda117/nompi/nvcpp/debug/shared:
+  extends:
+    - .build_and_test_template
+    - .default_variables
+    - .full_test_condition
+    - .use_gko_nvhpc227-cuda117-openmpi-gnu11-llvm14
+  variables:
+    CXX_COMPILER: nvc++
+    BUILD_OMP: 'ON'
+    BUILD_CUDA: 'ON'
+    BUILD_MPI: 'OFF'
+    BUILD_TYPE: Debug
+    FAST_TESTS: 'ON'
+    CXX_FLAGS: --diag_suppress=useless_using_declaration,declared_but_not_referenced
+    EXTRA_CMAKE_FLAGS: -DCMAKE_CUDA_FLAGS=-diag-suppress=177
+    _COMMENTS: '`--diag_suppress=useless_using_declaration,declared_but_not_referenced` disables common gflags warning. `-diag-suppress=188` disables spurious unused argument warning in CUDA.'
+
+build/cuda118/nompi/gcc/cuda/debug/shared:
+  extends:
+    - .build_and_test_tum_template
+    - .default_variables
+    - .full_test_condition
+    - .use_tum-nvidia
+  variables:
+    BUILD_CUDA: 'ON'
+    BUILD_HWLOC: 'OFF'
+    ENABLE_HALF: 'ON'
+    BUILD_TYPE: Debug
+    MODULE_LOAD: cmake/3.18.6 cuda/11.8.0 gcc/7.5.0
+
+build/cuda120/openmpi/gcc/cuda/release/static:
+  extends:
+    - .build_and_test_tum_template
+    - .default_variables
+    - .full_test_condition
+    - .use_tum-nvidia
+  variables:
+    BUILD_CUDA: 'ON'
+    BUILD_HWLOC: 'OFF'
+    ENABLE_HALF: 'ON'
+    BUILD_MPI: 'ON'
+    BUILD_SHARED_LIBS: 'OFF'
+    BUILD_TYPE: Release
+    MODULE_LOAD: cmake/3.18.6 cuda/12.0.1 gcc/12.4.0 openmpi/4.1.8
+
+build/cuda122/openmpi/gcc/cuda/release/shared:
+  extends:
+    - .build_and_test_tum_template
+    - .default_variables
+    - .full_test_condition
+    - .use_tum-nvidia
+  variables:
+    BUILD_CUDA: 'ON'
+    BUILD_HWLOC: 'OFF'
+    ENABLE_HALF: 'ON'
+    BUILD_TYPE: Release
+    MODULE_LOAD: cmake/3.20.6 cuda/12.2.2 gcc/12.4.0 openmpi/5.0.7
+
+build/cuda124/mpich/gcc/cuda/release/shared:
+  extends:
+    - .build_and_test_tum_template
+    - .default_variables
+    - .full_test_condition
+    - .use_tum-nvidia
+  variables:
+    BUILD_CUDA: 'ON'
+    BUILD_HWLOC: 'OFF'
+    BUILD_MPI: 'ON'
+    ENABLE_HALF: 'ON'
+    ENABLE_BFLOAT16: 'ON'
+    BUILD_TYPE: Release
+    MODULE_LOAD: cmake/3.30.8 cuda/12.6.3 gcc/13.3.0 mpich
+
+build/cuda126/nompi/gcc/cuda/release/shared:
+  extends:
+    - .build_and_test_tum_template
+    - .default_variables
+    - .full_test_condition
+    - .use_tum-nvidia
+  variables:
+    BUILD_CUDA: 'ON'
+    BUILD_HWLOC: 'OFF'
+    ENABLE_HALF: 'ON'
+    ENABLE_BFLOAT16: 'ON'
+    BUILD_TYPE: Release
+    MODULE_LOAD: cmake/3.30.8 cuda/12.6.3 gcc/13.3.0
+
+build/cuda130/nompi/gcc/cuda/release/shared:
+  extends:
+    - .build_and_test_tum_template
+    - .default_variables
+    - .full_test_condition
+    - .use_tum-nvidia
+  variables:
+    BUILD_CUDA: 'ON'
+    BUILD_HWLOC: 'OFF'
+    ENABLE_HALF: 'ON'
+    ENABLE_BFLOAT16: 'ON'
+    BUILD_TYPE: Release
+    MODULE_LOAD: cmake/3.30.8 cuda/13.0.2 gcc/14.3.0
+
+build/cuda131/nompi/gcc/cuda/release/shared:
+  extends:
+    - .build_and_test_tum_template
+    - .default_variables
+    - .full_test_condition
+    - .use_tum-nvidia
+  variables:
+    BUILD_CUDA: 'ON'
+    BUILD_HWLOC: 'OFF'
+    ENABLE_HALF: 'ON'
+    ENABLE_BFLOAT16: 'ON'
+    BUILD_TYPE: Release
+    MODULE_LOAD: cmake/3.30.8 cuda/13.1.1 gcc/14.3.0
+
+build/amd/openmpi/gcc/rocm620/debug/shared:
+  extends:
+    - .build_and_test_tum_template
+    - .default_variables
+    - .quick_test_condition
+    - .use_tum-amd
+  variables:
+    BUILD_HIP: 'ON'
+    BUILD_HWLOC: 'OFF'
+    BUILD_MPI: 'ON'
+    BUILD_OMP: 'ON'
+    BUILD_TYPE: Debug
+    BUILD_SHARED_LIBS: 'ON'
+    MODULE_LOAD: cmake/3.24.4 rocm/6.2.0 gcc/13.3.0 openmpi/5.0.7
+
+build/amd/openmpi/gcc/rocm634_wo_omp/release/shared:
+  extends:
+    - .build_and_test_tum_template
+    - .default_variables
+    - .full_test_condition
+    - .use_tum-amd
+  variables:
+    BUILD_HIP: 'ON'
+    BUILD_HWLOC: 'OFF'
+    BUILD_MPI: 'ON'
+    BUILD_TYPE: Release
+    MODULE_LOAD: cmake/3.29.6 rocm/6.3.4 gcc/13.3.0 openmpi/5.0.7
+    _COMMENTS: MI50 is officially supported by ROCm < 6.4.0
+
+build/amd/nompi/gcc/rocm644/debug/static:
+  extends:
+    - .build_and_test_tum_template
+    - .default_variables
+    - .full_test_condition
+    - .use_tum-amd-mi210
+  variables:
+    BUILD_HIP: 'ON'
+    BUILD_HWLOC: 'OFF'
+    BUILD_MPI: 'OFF'
+    BUILD_OMP: 'OFF'
+    BUILD_TYPE: Debug
+    BUILD_SHARED_LIBS: 'OFF'
+    MODULE_LOAD: cmake/3.29.6 rocm/6.4.4 gcc/14.3.0
+    _COMMENTS: MI50 is not officially supported by ROCm >= 6.4.0
+
+build/amd/nompi/gcc/rocm710/release/shared:
+  extends:
+    - .build_and_test_tum_template
+    - .default_variables
+    - .full_test_condition
+    - .use_tum-amd-mi210
+  variables:
+    BUILD_HIP: 'ON'
+    BUILD_HWLOC: 'OFF'
+    BUILD_MPI: 'OFF'
+    BUILD_OMP: 'OFF'
+    BUILD_TYPE: Release
+    MODULE_LOAD: cmake/3.29.6 rocm/7.1.0 gcc/14.3.0
+
+build/amd/nompi/gcc/rocm720/release/shared:
+  extends:
+    - .build_and_test_tum_template
+    - .default_variables
+    - .full_test_condition
+    - .use_tum-amd
+  variables:
+    BUILD_HIP: 'ON'
+    BUILD_OMP: 'OFF'
+    BUILD_HWLOC: 'OFF'
+    BUILD_TYPE: Release
+    ENABLE_BFLOAT16: 'OFF'
+    ENABLE_HALF: 'OFF'
+    MODULE_LOAD: cmake/3.25.3 rocm/7.2.0 gcc/12.4.0
+
+build/nocuda/nompi/gcc/core/debug/static:
+  extends:
+    - .build_and_test_template
+    - .default_variables
+    - .full_test_condition
+    - .use_gko-nocuda-nompi-gnu9-llvm8
+  variables:
+    BUILD_TYPE: Debug
+    FAST_TESTS: 'ON'
+    BUILD_SHARED_LIBS: 'OFF'
+    BUILD_HWLOC: 'OFF'
+
+build/nocuda/nompi/clang/core/release/shared:
+  extends:
+    - .build_and_test_template
+    - .default_variables
+    - .full_test_condition
+    - .use_gko-nocuda-nompi-gnu9-llvm8
+  variables:
+    CXX_COMPILER: clang++
+    BUILD_TYPE: Release
+    ENABLE_HALF: 'OFF'
+    ENABLE_BFLOAT16: 'ON'
+
+build/nocuda/nompi/gcc/omp/release/shared:
+  extends:
+    - .build_and_test_template
+    - .default_variables
+    - .quick_test_condition
+    - .use_gko-nocuda-nompi-gnu9-llvm8
+  variables:
+    BUILD_OMP: 'ON'
+    BUILD_TYPE: Release
+    ENABLE_HALF: 'ON'
+    ENABLE_BFLOAT16: 'ON'
+
+build/nocuda/openmpi/clang/omp/debug/static:
+  extends:
+    - .build_and_test_template
+    - .default_variables
+    - .full_test_condition
+    - .use_gko-nocuda-openmpi-gnu9-llvm8
+  variables:
+    CXX_COMPILER: clang++
+    BUILD_OMP: 'ON'
+    MPI_AS_ROOT: 'ON'
+    BUILD_MPI: 'ON'
+    BUILD_TYPE: Debug
+    FAST_TESTS: 'ON'
+    BUILD_SHARED_LIBS: 'OFF'
+
+build/nocuda/openmpi/clang/omp/glibcxx-debug-release/shared:
+  extends:
+    - .build_and_test_template
+    - .default_variables
+    - .full_test_condition
+    - .use_gko-nocuda-nompi-gnu9-llvm8
+  variables:
+    CXX_COMPILER: clang++
+    BUILD_OMP: 'ON'
+    MPI_AS_ROOT: 'ON'
+    BUILD_MPI: 'ON'
+    CXX_FLAGS: -Wpedantic -D_GLIBCXX_DEBUG=1
+    BUILD_TYPE: Release
+    _COMMENTS: We use Release here bacause glibcxx tests are prohibitively slow in Debug
+    _JOBNAME_SUFFIX: /glibcxx
+
+build/nocuda/nompi/gcc/omp/release/static:
+  extends:
+    - .build_and_test_template
+    - .default_variables
+    - .full_test_condition
+    - .use_gko-nocuda-nompi-gnu9-llvm8
+  variables:
+    BUILD_OMP: 'ON'
+    BUILD_TYPE: Release
+    BUILD_SHARED_LIBS: 'OFF'
+    ENABLE_HALF: 'OFF'
+    ENABLE_BFLOAT16: 'ON'
+
+build/nocuda-nomixed/nompi/clang/omp/release/static:
+  extends:
+    - .build_and_test_template
+    - .default_variables
+    - .full_test_condition
+    - .use_gko-nocuda-nompi-gnu9-llvm8
+  variables:
+    CXX_COMPILER: clang++
+    BUILD_OMP: 'ON'
+    BUILD_TYPE: Release
+    BUILD_SHARED_LIBS: 'OFF'
+    MIXED_PRECISION: 'OFF'
+
+build/nocuda-nomixed/openmpi/gcc/omp/release/shared:
+  extends:
+    - .build_and_test_template
+    - .default_variables
+    - .full_test_condition
+    - .use_gko-nocuda-nompi-gnu9-llvm8
+  variables:
+    BUILD_MPI: 'ON'
+    MPI_AS_ROOT: 'ON'
+    BUILD_OMP: 'ON'
+    BUILD_TYPE: Release
+    MIXED_PRECISION: 'OFF'
+
+build/nocuda-nomixed/nompi/clang/omp/debug/static:
+  extends:
+    - .build_and_test_template
+    - .default_variables
+    - .full_test_condition
+    - .use_gko-nocuda-nompi-gnu9-llvm8
+  variables:
+    CXX_COMPILER: clang++
+    BUILD_OMP: 'ON'
+    BUILD_TYPE: Debug
+    BUILD_SHARED_LIBS: 'OFF'
+    MIXED_PRECISION: 'OFF'
+
+build/nogpu/nompi/gcc/omp/debug/static:
+  extends:
+    - .build_and_test_tum_template
+    - .default_variables
+    - .full_test_condition
+    - .use_tum-cpu
+  variables:
+    BUILD_OMP: 'ON'
+    BUILD_TYPE: Debug
+    BUILD_SHARED_LIBS: 'OFF'
+    MIXED_PRECISION: 'OFF'
+    BUILD_HWLOC: 'OFF'
+    MODULE_LOAD: cmake/3.18.6 gcc/7.5.0
+
+build/nogpu/nompi/gcc/omp/release/static:
+  extends:
+    - .build_and_test_tum_template
+    - .default_variables
+    - .full_test_condition
+    - .use_tum-cpu
+  variables:
+    BUILD_OMP: 'ON'
+    BUILD_TYPE: Release
+    BUILD_SHARED_LIBS: 'OFF'
+    ENABLE_HALF: 'ON'
+    ENABLE_BFLOAT16: 'ON'
+    BUILD_HWLOC: 'OFF'
+    MODULE_LOAD: cmake/3.18.6 gcc/9.5.0
+
+build/nogpu/nompi/gcc/noomp/release/shared:
+  extends:
+    - .build_and_test_tum_template
+    - .default_variables
+    - .full_test_condition
+    - .use_tum-cpu
+  variables:
+    BUILD_TYPE: Release
+    BUILD_HWLOC: 'OFF'
+    MODULE_LOAD: cmake/3.18.6 gcc/9.5.0
+
+build/nogpu/nompi/gcc12/noomp/release/shared:
+  extends:
+    - .build_and_test_tum_template
+    - .default_variables
+    - .full_test_condition
+    - .use_tum-cpu
+  variables:
+    BUILD_TYPE: Release
+    BUILD_HWLOC: 'OFF'
+    MODULE_LOAD: cmake/3.18.6 gcc/12.4.0
+
+build/nogpu/mpi/gcc13/noomp/release/shared:
+  extends:
+    - .build_and_test_tum_template
+    - .default_variables
+    - .quick_test_condition
+    - .use_tum-cpu
+  variables:
+    BUILD_TYPE: Release
+    BUILD_HWLOC: 'OFF'
+    BUILD_MPI: 'ON'
+    MODULE_LOAD: cmake/3.18.6 gcc/13.3.0 openmpi
+
+build/icpx20231/gpu/release/shared:
+  extends:
+    - .build_and_test_template
+    - .default_variables
+    - .quick_test_condition
+    - .use_gko-oneapi20231-gpu
+  variables:
+    CXX_COMPILER: icpx
+    CXX_FLAGS: -Wpedantic -ffp-model=precise
+    BUILD_SYCL: 'ON'
+    BUILD_TYPE: Release
+    BUILD_SHARED_LIBS: 'ON'
+    DPCPP_SINGLE_MODE: 'ON'
+    ONEAPI_DEVICE_SELECTOR: opencl:gpu
+    BUILD_HWLOC: 'OFF'
+
+build/dpcpp20231/gpu/release/shared:
+  extends:
+    - .build_and_test_tum_template
+    - .default_variables
+    - .full_test_condition
+    - .use_tum-intel-fp32
+  variables:
+    CXX_COMPILER: dpcpp
+    CXX_FLAGS: -Wpedantic -ffp-model=precise
+    BUILD_SYCL: 'ON'
+    BUILD_TYPE: Release
+    BUILD_SHARED_LIBS: 'ON'
+    DPCPP_SINGLE_MODE: 'ON'
+    MODULE_LOAD: cmake/3.21.7 intel-oneapi-compilers/2023.1.0 intel-oneapi-dpl/2022.1.0 intel-oneapi-tbb/2021.9.0 intel-oneapi-mkl/2023.1.0
+    ONEAPI_DEVICE_SELECTOR: opencl:gpu
+    BUILD_HWLOC: 'OFF'
+
+build/icpx202402/gpu/release/shared:
+  extends:
+    - .build_and_test_tum_template
+    - .default_variables
+    - .full_test_condition
+    - .use_tum-intel
+  variables:
+    CXX_COMPILER: icpx
+    CXX_FLAGS: -Wpedantic -ffp-model=precise -Wno-deprecated-declarations
+    BUILD_SYCL: 'ON'
+    BUILD_HWLOC: 'OFF'
+    BUILD_TYPE: Release
+    ENABLE_HALF: 'OFF'
+    ENABLE_BFLOAT16: 'OFF'
+    ONEAPI_DEVICE_SELECTOR: opencl:gpu
+    MODULE_LOAD: cmake/3.21.7 intel-oneapi-compilers/2024.0.2 intel-oneapi-dpl/2022.3.0 intel-oneapi-tbb/2021.11.0 intel-oneapi-mkl/2024.0.0
+
+build/icpx202421/gpu/release/shared:
+  extends:
+    - .build_and_test_tum_template
+    - .default_variables
+    - .full_test_condition
+    - .use_tum-intel
+  variables:
+    CXX_COMPILER: icpx
+    CXX_FLAGS: -Wpedantic -ffp-model=precise -Wno-deprecated-declarations
+    BUILD_SYCL: 'ON'
+    BUILD_OMP: 'ON'
+    BUILD_HWLOC: 'OFF'
+    BUILD_TYPE: Release
+    ENABLE_HALF: 'ON'
+    ENABLE_BFLOAT16: 'ON'
+    ONEAPI_DEVICE_SELECTOR: level_zero:gpu
+    MODULE_LOAD: cmake/3.28.6 intel-oneapi-compilers/2024.2.1 intel-oneapi-dpl/2022.6.1 intel-oneapi-tbb/2021.13.1 intel-oneapi-mkl/2024.2.1
+
+build/dpcpp/gpu/release/shared:
+  extends:
+    - .build_and_test_template
+    - .default_variables
+    - .full_test_condition
+    - .use_gko-oneapi-gpu
+  variables:
+    CXX_COMPILER: dpcpp
+    CXX_FLAGS: -Wpedantic -ffp-model=precise
+    BUILD_SYCL: 'ON'
+    BUILD_TYPE: Release
+    BUILD_SHARED_LIBS: 'ON'
+    DPCPP_SINGLE_MODE: 'ON'
+    ONEAPI_DEVICE_SELECTOR: '*:gpu'
+    BUILD_HWLOC: 'OFF'
+    _COMMENTS: This gives two available backends of GPU on tests
diff --git a/.gitlab/image.yml b/.gitlab/image.yml
index a609a26fdb5..f80d9b074a4 100644
--- a/.gitlab/image.yml
+++ b/.gitlab/image.yml
@@ -39,12 +39,6 @@
     - nvidia-gpus-p100
     - tum
 
-.use_gko-rocm45-nompi-gnu8-llvm8:
-  image: ginkgohub/rocm:45-mvapich2-gnu8-llvm8
-  tags:
-    - private_ci
-    - amd-gpu
-
 .use_gko-rocm514-nompi-gnu11-llvm11:
   image: ginkgohub/rocm:514-openmpi-gnu11-llvm11
   tags:
diff --git a/.gitlab/scripts.yml b/.gitlab/scripts.yml
index 8a187f85bc0..7e5b1a65bb0 100644
--- a/.gitlab/scripts.yml
+++ b/.gitlab/scripts.yml
@@ -54,7 +54,7 @@
         -DGINKGO_BUILD_HIP=${BUILD_HIP} -DGINKGO_BUILD_SYCL=${BUILD_SYCL}
         -DGINKGO_BUILD_MPI=${BUILD_MPI} ${MPI_STR}
         -DGINKGO_BUILD_HWLOC=${BUILD_HWLOC}
-        -DGINKGO_BUILD_PAPI_SDE=${BUILD_PAPI_SDE}
+        -DGINKGO_BUILD_PAPI_SDE=OFF
         -DGINKGO_BUILD_TESTS=ON -DGINKGO_BUILD_EXAMPLES=ON
         -DGINKGO_FAST_TESTS=${FAST_TESTS}
         -DGINKGO_TEST_NONDEFAULT_STREAM=${NONDEFAULT_STREAM}
@@ -78,8 +78,6 @@
     - if [ -n "${CUDA_ARCH}" ]; then
       export CUDA_ARCH_STR=-DGINKGO_CUDA_ARCHITECTURES=${CUDA_ARCH};
       fi
-    - if [ -n "${SYCL_DEVICE_TYPE}" ]; then export SYCL_DEVICE_TYPE; fi
-    - if [ -n "${SYCL_DEVICE_FILTER}" ]; then export SYCL_DEVICE_FILTER; fi
     - if [ -n "${ONEAPI_DEVICE_SELECTOR}" ]; then export ONEAPI_DEVICE_SELECTOR; fi
     - if [[ "${MPI_AS_ROOT}" == "ON" ]];then
       export OMPI_ALLOW_RUN_AS_ROOT=1;
@@ -99,13 +97,13 @@
         -DGINKGO_BUILD_HIP=${BUILD_HIP} -DGINKGO_BUILD_SYCL=${BUILD_SYCL}
         -DGINKGO_BUILD_MPI=${BUILD_MPI} ${MPI_STR}
         -DGINKGO_BUILD_HWLOC=${BUILD_HWLOC}
-        -DGINKGO_BUILD_PAPI_SDE=${BUILD_PAPI_SDE}
+        -DGINKGO_BUILD_PAPI_SDE=OFF
         -DGINKGO_BUILD_TESTS=ON -DGINKGO_BUILD_EXAMPLES=ON
         -DGINKGO_FAST_TESTS=${FAST_TESTS}
         -DGINKGO_ENABLE_HALF=${ENABLE_HALF}
         -DGINKGO_ENABLE_BFLOAT16=${ENABLE_BFLOAT16}
         -DGINKGO_MIXED_PRECISION=${MIXED_PRECISION}
-        -DGINKGO_CONFIG_LOG_DETAILED=${CONFIG_LOG}
+        -DGINKGO_CONFIG_LOG_DETAILED=ON
         -DGINKGO_DPCPP_SINGLE_MODE=${DPCPP_SINGLE_MODE}
     - ninja -j${NUM_CORES} -l${CI_LOAD_LIMIT} install
     - awk '!/^#/ { print ($2 - $1)/1000 " " $4 }' .ninja_log | sort -nr
@@ -117,8 +115,6 @@
     - pushd test/test_install
     - ninja install
     - popd
-    - if [ -n "${SYCL_DEVICE_TYPE}" ]; then unset SYCL_DEVICE_TYPE; fi
-    - if [ -n "${SYCL_DEVICE_FILTER}" ]; then unset SYCL_DEVICE_FILTER; fi
     - if [[ "${BUILD_SHARED_LIBS}" == "ON" || "${BUILD_HIP}" != "ON" ]]; then
       PKG_CONFIG_PATH=${INSTALL_PREFIX}/lib/pkgconfig:$PKG_CONFIG_PATH LD_LIBRARY_PATH=${INSTALL_PREFIX}/lib:$LD_LIBRARY_PATH ninja test_pkgconfig;
       fi
@@ -139,8 +135,6 @@
     - if [ -n "${CUDA_ARCH}" ]; then
       export CUDA_ARCH_STR=-DGINKGO_CUDA_ARCHITECTURES=${CUDA_ARCH};
       fi
-    - if [ -n "${SYCL_DEVICE_TYPE}" ]; then export SYCL_DEVICE_TYPE; fi
-    - if [ -n "${SYCL_DEVICE_FILTER}" ]; then export SYCL_DEVICE_FILTER; fi
     - if [ -n "${ONEAPI_DEVICE_SELECTOR}" ]; then export ONEAPI_DEVICE_SELECTOR; fi
     - if [[ "${MPI_AS_ROOT}" == "ON" ]];then
       export OMPI_ALLOW_RUN_AS_ROOT=1;
@@ -180,7 +174,7 @@
         -DGINKGO_BUILD_HIP=${BUILD_HIP} -DGINKGO_BUILD_SYCL=${BUILD_SYCL}
         -DGINKGO_BUILD_MPI=${BUILD_MPI} ${MPI_STR}
         -DGINKGO_BUILD_HWLOC=${BUILD_HWLOC}
-        -DGINKGO_BUILD_PAPI_SDE=${BUILD_PAPI_SDE}
+        -DGINKGO_BUILD_PAPI_SDE=OFF
         -DGINKGO_BUILD_TESTS=ON -DGINKGO_BUILD_EXAMPLES=ON
         -DGINKGO_FAST_TESTS=${FAST_TESTS}
         -DGINKGO_ENABLE_HALF=${ENABLE_HALF}
@@ -207,102 +201,7 @@
     - pushd test/test_install
     - ninja install
     - popd
-    - if [ -n "${SYCL_DEVICE_TYPE}" ]; then unset SYCL_DEVICE_TYPE; fi
-    - if [ -n "${SYCL_DEVICE_FILTER}" ]; then unset SYCL_DEVICE_FILTER; fi
     - if [[ "${BUILD_SHARED_LIBS}" == "ON" || "${BUILD_HIP}" != "ON" ]]; then
       PKG_CONFIG_PATH=${INSTALL_PREFIX}/lib64/pkgconfig:$PKG_CONFIG_PATH LD_LIBRARY_PATH=${INSTALL_PREFIX}/lib64:$LD_LIBRARY_PATH ninja test_pkgconfig;
       fi
   dependencies: []
-
-
-.horeka_test_template:
-  stage: test
-  before_script:
-    # Sanity checks
-    - if [ -z "${USE_NAME}" ]; then exit 111; fi
-    - if [ -z "${SLURM_PARTITION}" ]; then exit 222; fi
-    - if [[ ! "${SLURM_GRES}" =~ "^gpu*"  ]]; then export NVIDIA_VISIBLE_DEVICES=void; fi
-  variables:
-    GIT_CLEAN_FLAGS: none
-  script:
-    # We should be reusing the previous image and build directory.
-    # Speedup the tests by limiting the number of OMP threads to CI parallelism.
-    - export OMP_NUM_THREADS=${CI_PARALLELISM}
-    # The test job should have the name as the image job, only changing
-    # build to test
-    - cd ${CI_PROJECT_DIR}${CI_PROJECT_DIR_SUFFIX}
-    - cd ${CI_JOB_NAME/test/build}
-    - export INSTALL_PREFIX=`pwd`/install_prefix
-    - |
-        (( $(ctest -N | tail -1 | sed 's/Total Tests: //') != 0 )) || exit 1
-    - ctest --output-on-failure --timeout 6000 ${CTEST_EXTRA_ARGS}
-    - ninja test_exportbuild
-    - ninja test_install
-    - pushd test/test_install
-    - ninja install
-    - popd
-    - if [[ "${BUILD_SHARED_LIBS}" == "ON" || "${BUILD_HIP}" != "ON" ]]; then
-      PKG_CONFIG_PATH=${INSTALL_PREFIX}/lib/pkgconfig:$PKG_CONFIG_PATH LD_LIBRARY_PATH=${INSTALL_PREFIX}/lib:$LD_LIBRARY_PATH ninja test_pkgconfig;
-      fi
-  cache: []
-
-
-.horeka_benchmark_before_script_template:
-  before_script:
-    # Sanity checks
-    - if [ -z "${USE_NAME}" ]; then exit 111; fi
-    - if [ -z "${SLURM_PARTITION}" ]; then exit 222; fi
-    - if [[ ! "${SLURM_GRES}" =~ "^gpu*"  ]]; then export NVIDIA_VISIBLE_DEVICES=void; fi
-    - if [[ -z "${EXECUTOR}" ]]; then exit 233; fi
-    - if [[ -z "${SYSTEM_NAME}" ]]; then exit 233; fi
-    - if [[ -z "${BENCHMARK}" ]]; then exit 233; fi
-    # Set environment variables
-    - eval $(ssh-agent -s)
-    - echo "${BOT_KEY}" | tr -d '\r' | ssh-add - >/dev/null
-    - mkdir -p ~/.ssh
-    - chmod 700 ~/.ssh
-    - ssh-keyscan -t rsa github.com >> ~/.ssh/known_hosts
-    - chmod 644 ~/.ssh/known_hosts
-    - echo -e "Host *\n\tStrictHostKeyChecking no\n\n" > ~/.ssh/config
-    - unset OMP_NUM_THREADS
-    - git config --global user.name "${BOT_USER}"
-    - git config --global user.email "${BOT_EMAIL}"
-
-
-.benchmark_template:
-  extends:
-    - .horeka_benchmark_before_script_template
-  variables:
-    GIT_CLEAN_FLAGS: none
-  script:
-    # Install rsync
-    - pushd /
-    - git clone https://github.com/WayneD/rsync -b v3.2.4
-    - pushd rsync
-    - ./configure --disable-lz4 --disable-openssl --disable-doc --disable-xxhash --disable-md2man
-    - make -j10 install
-    - popd
-    - popd
-    # Setup ssget
-    - git clone https://github.com/ginkgo-project/ssget
-    - sed -i 's:ARCHIVE_LOCATION="${HOME}/.config/ssget":ARCHIVE_LOCATION="/ssget":' ssget/ssget
-    - export PATH=$PWD/ssget:$PATH
-    # Setup the benchmark
-    - cd ${CI_PROJECT_DIR}${CI_PROJECT_DIR_SUFFIX}
-    - cd ${CI_JOB_NAME}-build
-    - chmod +x benchmark/run_all_benchmarks.sh
-    - cd benchmark
-    - git clone ${BENCHMARK_REPO} data-repo
-    # Use the representative matrix list and launch benchmark
-    - export MATRIX_LIST_FILE=$PWD/data-repo/data/represent.list
-    - ./run_all_benchmarks.sh
-    # Publish the results
-    - rsync -av results/ data-repo/data/
-    - cd data-repo/data/
-    - ./build-list . > list.json
-    - ./agregate < list.json > agregate.json
-    - chmod +x represent
-    - ./represent $PWD > represent.json
-    - git add -A
-    - git diff --quiet HEAD || (git commit -m "Benchmark ${BENCHMARK} on ${EXECUTOR} with ${SYSTEM_NAME} of ginkgo-project/ginkgo@${CI_COMMIT_SHORT_SHA}" && git push)
-  cache: []
diff --git a/.gitlab/variables.yml b/.gitlab/variables.yml
index 074322249b5..99e4768900b 100644
--- a/.gitlab/variables.yml
+++ b/.gitlab/variables.yml
@@ -1,7 +1,6 @@
 # Templates with reasonable defaults for builds and tests
 .default_variables:
   variables:
-    BENCHMARK_SERVER: "FINECI"
     CXX_COMPILER: "g++"
     CUDA_COMPILER: "nvcc"
     BUILD_TYPE: "Debug"
@@ -12,7 +11,6 @@
     BUILD_HIP: "OFF"
     BUILD_SYCL: "OFF"
     BUILD_HWLOC: "ON"
-    BUILD_PAPI_SDE: "OFF"
     BUILD_MPI: "OFF"
     MPI_AS_ROOT: "OFF"
     FAST_TESTS: "OFF"
@@ -21,7 +19,12 @@
     MIXED_PRECISION: "ON"
     ENABLE_HALF: "ON"
     ENABLE_BFLOAT16: "OFF"
-    CONFIG_LOG: "ON"
     CXX_FLAGS: "-Wpedantic"
     EXTRA_CMAKE_FLAGS: ""
     CI_PROJECT_DIR_SUFFIX: ""
+    CUDA_ARCH: ""
+    MDOULE_LOAD: ""
+    SPACK_LOAD: ""
+    ONEAPI_DEVICE_SELECTOR: ""
+    _JOBNAME_SUFFIX: ""
+    _COMMENTS: ""
diff --git a/CMakeLists.txt b/CMakeLists.txt
index 088a1e4cb40..dcfd48dc3e3 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -443,7 +443,7 @@ endif()
 
 # Try to find the third party packages before using our subdirectories
 if(GINKGO_BUILD_TESTS)
-    find_package(GTest 1.10.0) # No need for QUIET as CMake ships FindGTest
+    find_package(GTest 1.12.0) # No need for QUIET as CMake ships FindGTest
 endif()
 if(GINKGO_BUILD_BENCHMARKS)
     find_package(gflags 2.2.2 QUIET)
diff --git a/dpcpp/get_info.cmake b/dpcpp/get_info.cmake
index ee9c0398f3e..ebf4f132546 100644
--- a/dpcpp/get_info.cmake
+++ b/dpcpp/get_info.cmake
@@ -3,5 +3,5 @@ ginkgo_print_module_footer(${detailed_log} "DPCPP variables:")
 ginkgo_print_variable(${detailed_log} "GINKGO_DPCPP_FLAGS")
 ginkgo_print_variable(${detailed_log} "GINKGO_DPCPP_SINGLE_MODE")
 ginkgo_print_module_footer(${detailed_log} "DPCPP environment variables:")
-ginkgo_print_env_variable(${detailed_log} "SYCL_DEVICE_FILTER")
+ginkgo_print_env_variable(${detailed_log} "ONEAPI_DEVICE_SELECTOR")
 ginkgo_print_module_footer(${detailed_log} "")