NVIDIA · alliepiper · May 13, 2026 · May 13, 2026 · May 13, 2026
diff --git a/.github/actions/compute-matrix/compute-matrix.sh b/.github/actions/compute-matrix/compute-matrix.sh
@@ -18,6 +18,20 @@ extract_matrix() {
   local per_cuda_compiler_matrix="$(echo "$nvcc_full_matrix" | jq -cr ' group_by(.cuda + .compiler.name) | map({(.[0].cuda + "-" + .[0].compiler.name): .}) | add')"
   write_output "PER_CUDA_COMPILER_MATRIX"  "$per_cuda_compiler_matrix"
   write_output "PER_CUDA_COMPILER_KEYS" "$(echo "$per_cuda_compiler_matrix" | jq -r 'keys | @json')"
+
+  local windows_matrix="$(echo "$matrix" | jq -cr '
+    (.windows // [])
+    | map({
+        cuda: .cuda,
+        host: (.host // (.compiler.name + (.compiler.version | tostring))),
+        cpu: (.cpu // "amd64"),
+        runner: (.runner // ("windows-" + (.cpu // "amd64") + "-cpu16")),
+        std: ((.std // "17") | tostring),
+        arch: (.arch // ""),
+        image: (.image // "")
+      })
+  ')"
+  write_output "WINDOWS_MATRIX" "$windows_matrix"
 }
 
 main() {

diff --git a/.github/workflows/build-windows.yml b/.github/workflows/build-windows.yml
@@ -0,0 +1,170 @@
+name: build windows
+
+defaults:
+  run:
+    shell: powershell
+
+on:
+  workflow_call:
+    inputs:
+      cuda:
+        type: string
+        required: false
+        default: "13.0"
+      host:
+        type: string
+        required: false
+        default: "cl14.44"
+      devcontainer_version:
+        type: string
+        required: false
+        default: "25.12"
+      image:
+        type: string
+        required: false
+        default: ""
+      runner:
+        type: string
+        required: false
+        default: "windows-amd64-cpu16"
+      std:
+        type: string
+        required: false
+        default: "17"
+      arch:
+        type: string
+        required: false
+        default: ""
+  workflow_dispatch:
+    inputs:
+      cuda:
+        description: "CUDA Toolkit version"
+        type: string
+        required: false
+        default: "13.0"
+      host:
+        description: "Host compiler tag"
+        type: string
+        required: false
+        default: "cl14.44"
+      devcontainer_version:
+        description: "RAPIDS devcontainer version"
+        type: string
+        required: false
+        default: "25.12"
+      image:
+        description: "Windows devcontainer image override"
+        type: string
+        required: false
+        default: ""
+      runner:
+        description: "Windows runner"
+        type: string
+        required: false
+        default: "windows-amd64-cpu16"
+      std:
+        description: "C++/CUDA standard"
+        type: choice
+        options: ["17", "20"]
+        required: false
+        default: "17"
+      arch:
+        description: "CMAKE_CUDA_ARCHITECTURES override"
+        type: string
+        required: false
+        default: ""
+
+permissions:
+  contents: read
+
+jobs:
+  build-windows:
+    name: Build Windows CUDA${{ inputs.cuda }} ${{ inputs.host }} C++${{ inputs.std }}
+    runs-on: ${{ github.repository == 'NVIDIA/nvbench' && inputs.runner || 'windows-latest' }}
+    permissions:
+      id-token: write
+      contents: read
+    env:
+      WINDOWS_CI_IMAGE: ${{ inputs.image != '' && inputs.image || format('rapidsai/devcontainers:{0}-cuda{1}-{2}', inputs.devcontainer_version, inputs.cuda, inputs.host) }}
+      SCCACHE_BUCKET: rapids-sccache-devs
+      SCCACHE_REGION: us-east-2
+      SCCACHE_IDLE_TIMEOUT: "0"
+      SCCACHE_S3_USE_SSL: "true"
+      SCCACHE_S3_NO_CREDENTIALS: "false"
+      SCCACHE_S3_USE_PREPROCESSOR_CACHE_MODE: "true"
+      SCCACHE_S3_PREPROCESSOR_CACHE_KEY_PREFIX: nvbench-preprocessor-cache
+    steps:
+      - name: Checkout repo
+        uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
+          fetch-tags: true
+          persist-credentials: false
+
+      - name: Add NVCC problem matcher
+        run: |
+          echo "::add-matcher::.github/problem-matchers/problem-matcher.json"
+
+      - name: Get AWS credentials for sccache bucket
+        if: ${{ github.repository == 'NVIDIA/nvbench' }}
+        uses: aws-actions/configure-aws-credentials@v4
+        with:
+          role-to-assume: arn:aws:iam::279114543810:role/gha-oidc-NVIDIA
+          aws-region: us-east-2
+          role-duration-seconds: 43200
+
+      - name: Fetch Windows devcontainer image
+        run: |
+          docker pull "$env:WINDOWS_CI_IMAGE"
+
+      - name: Build NVBench
+        env:
+          NVBENCH_WINDOWS_STD: ${{ inputs.std }}
+          NVBENCH_WINDOWS_ARCH: ${{ inputs.arch }}
+        run: |
+          $ErrorActionPreference = "Stop"
+
+          $hostRepo = "${{ github.workspace }}".Replace('\', '/')
+          $containerRepo = "C:/nvbench"
+          $script = Join-Path $env:RUNNER_TEMP "nvbench-windows-ci.ps1"
+
+          @"
+          `$ErrorActionPreference = 'Stop'
+          git config --global --add safe.directory '$containerRepo'
+          & '$containerRepo/ci/windows/build_nvbench.ps1' -std '$env:NVBENCH_WINDOWS_STD' -arch '$env:NVBENCH_WINDOWS_ARCH'
+          exit `$LASTEXITCODE
+          "@ | Set-Content -Path $script -Encoding UTF8
+
+          $containerScript = "$containerRepo/nvbench-windows-ci.ps1"
+          Copy-Item $script "${{ github.workspace }}\nvbench-windows-ci.ps1"
+
+          $dockerArgs = @(
+            "run", "--rm",
+            "--mount", "type=bind,source=$hostRepo,target=$containerRepo",
+            "--workdir", $containerRepo,
+            "--isolation=process",
+            "--env", "AWS_ACCESS_KEY_ID=$env:AWS_ACCESS_KEY_ID",
+            "--env", "AWS_SECRET_ACCESS_KEY=$env:AWS_SECRET_ACCESS_KEY",
+            "--env", "AWS_SESSION_TOKEN=$env:AWS_SESSION_TOKEN",
+            "--env", "CI=true",
+            "--env", "GITHUB_ACTIONS=$env:GITHUB_ACTIONS",
+            "--env", "GITHUB_REF_NAME=$env:GITHUB_REF_NAME",
+            "--env", "GITHUB_REPOSITORY=$env:GITHUB_REPOSITORY",
+            "--env", "GITHUB_RUN_ID=$env:GITHUB_RUN_ID",
+            "--env", "GITHUB_SHA=$env:GITHUB_SHA",
+            "--env", "SCCACHE_BUCKET=$env:SCCACHE_BUCKET",
+            "--env", "SCCACHE_IDLE_TIMEOUT=$env:SCCACHE_IDLE_TIMEOUT",
+            "--env", "SCCACHE_REGION=$env:SCCACHE_REGION",
+            "--env", "SCCACHE_S3_NO_CREDENTIALS=$env:SCCACHE_S3_NO_CREDENTIALS",
+            "--env", "SCCACHE_S3_PREPROCESSOR_CACHE_KEY_PREFIX=$env:SCCACHE_S3_PREPROCESSOR_CACHE_KEY_PREFIX",
+            "--env", "SCCACHE_S3_USE_PREPROCESSOR_CACHE_MODE=$env:SCCACHE_S3_USE_PREPROCESSOR_CACHE_MODE",
+            "--env", "SCCACHE_S3_USE_SSL=$env:SCCACHE_S3_USE_SSL",
+            "$env:WINDOWS_CI_IMAGE",
+            "powershell", "-NoLogo", "-NoProfile", "-ExecutionPolicy", "Bypass",
+            "-File", $containerScript
+          )
+
+          docker @dockerArgs
+          if ($LASTEXITCODE -ne 0) {
+            exit $LASTEXITCODE
+          }
diff --git a/.github/workflows/pr.yml b/.github/workflows/pr.yml
@@ -42,6 +42,7 @@ jobs:
       DEVCONTAINER_VERSION: ${{steps.set-outputs.outputs.DEVCONTAINER_VERSION}}
       PER_CUDA_COMPILER_MATRIX: ${{steps.set-outputs.outputs.PER_CUDA_COMPILER_MATRIX}}
       PER_CUDA_COMPILER_KEYS: ${{steps.set-outputs.outputs.PER_CUDA_COMPILER_KEYS}}
+      WINDOWS_MATRIX: ${{steps.set-outputs.outputs.WINDOWS_MATRIX}}
       base_sha: ${{ steps.export-pr-info.outputs.base_sha }}
       pr_number: ${{ steps.export-pr-info.outputs.pr_number }}
     steps:
@@ -76,6 +77,28 @@ jobs:
       per_cuda_compiler_matrix: ${{ toJSON(fromJSON(needs.compute-matrix.outputs.PER_CUDA_COMPILER_MATRIX)[ matrix.cuda_host_combination ]) }}
       devcontainer_version: ${{ needs.compute-matrix.outputs.DEVCONTAINER_VERSION }}
 
+  nvbench-windows:
+    name: NVBench Windows CUDA${{ matrix.config.cuda }} ${{ matrix.config.host }} C++${{ matrix.config.std }}
+    # TODO: Re-enable after https://github.com/NVIDIA/nvbench/pull/354 fixes the Windows build.
+    if: false
+    permissions:
+      id-token: write
+      contents: read
+    needs: compute-matrix
+    uses: ./.github/workflows/build-windows.yml
+    strategy:
+      fail-fast: false
+      matrix:
+        config: ${{ fromJSON(needs.compute-matrix.outputs.WINDOWS_MATRIX) }}
+    with:
+      cuda: ${{ matrix.config.cuda }}
+      host: ${{ matrix.config.host }}
+      runner: ${{ matrix.config.runner }}
+      std: ${{ matrix.config.std }}
+      arch: ${{ matrix.config.arch }}
+      image: ${{ matrix.config.image }}
+      devcontainer_version: ${{ needs.compute-matrix.outputs.DEVCONTAINER_VERSION }}
+
   python-wheels:
     name: Python Wheels
     permissions:
@@ -103,6 +126,7 @@ jobs:
     if: ${{ always() }} # need to use always() instead of !cancelled() because skipped jobs count as success
     needs:
       - nvbench
+      - nvbench-windows
       - python-wheels
       - verify-devcontainers
     steps:

diff --git a/ci/matrix.yaml b/ci/matrix.yaml
@@ -24,6 +24,9 @@ llvm18: &llvm18 { name: 'llvm', version: '18', exe: 'clang++' }
 llvm19: &llvm19 { name: 'llvm', version: '19', exe: 'clang++' }
 llvm20: &llvm20 { name: 'llvm', version: '20', exe: 'clang++' }
 
+# MSVC compiler configurations
+msvc2022: &msvc2022 { name: 'cl', version: '14.44', exe: 'cl' }
+
 # Each environment below will generate a unique build/test job
 # See the "compute-matrix" job in the workflow for how this is parsed and used
 # cuda: The CUDA Toolkit version
@@ -68,6 +71,8 @@ pull_request:
     - {cuda: *cuda_curr_max, compiler: *llvm18, cpu: 'amd64'}
     - {cuda: *cuda_curr_max, compiler: *llvm19, cpu: 'amd64'}
     - {cuda: *cuda_curr_max, compiler: *llvm20, cpu: 'amd64'}
+  windows:
+    - {cuda: *cuda_curr_max, compiler: *msvc2022, cpu: 'amd64', std: '17'}
 
 # Python wheel builds
 python_wheels: