diff --git a/programs/sbd/README.md b/programs/sbd/README.md
new file mode 100644
index 0000000..60c44c3
--- /dev/null
+++ b/programs/sbd/README.md
@@ -0,0 +1,85 @@
+# SBD benchmark
+
+This BenchKit application runs the NVIDIA Thrust TPB selected-basis
+diagonalization benchmark from `github.com/r-ccs-cms/sbd`.
+
+The RIKYU recipe uses the H2O cc-pVDZ FCIDUMP with the `1em7` selected alpha
+determinant file (about 628 million product determinants), one MPI rank per
+B200 GPU, 32 OpenMP threads per rank, and the rank-distributed/index-reordered/NCCL
+build configuration validated in the SubWG2 benchmark study. 32 threads is
+Rikyu's site-enforced CPU cap per requested GPU on shared nodes; it also
+covers the pre-Davidson setup pass (`RemakeHelpers`/`TaskCostSize`), which
+runs on the host rather than the GPU and is slow to the point of resembling
+a hang on large inputs if starved of threads. The inputs are staged from the
+SBD source clone into `artifacts/` by `build.sh` (see
+`data/h2o/README.md` for provenance) and can be overridden with
+`BK_SBD_INPUT_DIR` when a project-storage copy is preferred.
+
+The DGX Spark route uses the H2O `1em5` input (about 30.4 million product
+determinants), the Thrust backend with `cc120`, NVHPC/HPC-X CUDA 13 `26.3`,
+and one MPI rank on the single GB10 GPU. The larger `1em7` case reaches about
+98 GB before its first Davidson iteration and is killed by the 128 GB node's
+memory limit. `1em5` retains a substantial GPU workload while leaving enough
+memory headroom and runtime margin for continuous benchmarking. NCCL is disabled
+for the single-GPU case, and all three explicit communicator sizes are one. The
+Rikyu rows retain the `1em7` input and validated `1 x 2 x 2` explicit layout;
+their remaining rank factor becomes SBD's implicit Hamiltonian communicator.
+
+The R-CCS Cloud GH200 route uses the H2O `1em6` input (about 191 million
+product determinants), the Thrust backend with `cc90`, NVHPC/HPC-X CUDA 13
+`26.3`, and one MPI rank on its unified Grace Hopper superchip. No Slurm GPU
+request is made because `qc-gh200` exposes its single GPU as part of the node.
+NCCL is disabled for the single-rank case.
+
+The R-CCS Cloud FX700 route uses the H2O `1em4` input (about 2.38 million
+product determinants) with four MPI ranks and 12 OpenMP threads per rank,
+bound one rank per A64FX NUMA/CMG domain. SBD requires C++17 features missing
+from Fujitsu compiler 4.11.1's bundled libc++, so this route uses the system
+GCC 8.5/MPICH stack with 512-bit SVE enabled. It links Fujitsu's optimized
+LAPACK and its required runtime libraries by absolute path. The absolute paths
+are intentional: adding the Fujitsu library directory with `-L` causes
+MPICH's trailing `-lmpi` to resolve to Fujitsu MPI instead, mixing two MPI
+implementations in one executable.
+
+The input files are not committed to this repository; `build.sh` stages them
+from its `bk_fetch_source` clone of SBD's `data/h2o` directory into `artifacts/`.
+`data/h2o/README.md` records the upstream revision, checksums, and Apache-2.0
+provenance.
+
+The emitted FOM is SBD's internal Davidson time in seconds. Shell wall time is
+not used because MPI startup and scheduler overhead are substantial on RIKYU.
+The run is accepted only when its energy matches the established reference for
+its selected input within a combined `1e-12 + 1e-11 |E|` tolerance.
+
+The DGX Spark route was validated on `ng-dgx-m2` with NVHPC 26.3. The exact
+energy was `-76.24373504205295 Ha`; the internal Davidson FOM was
+`228.928853 s`, and the complete Slurm job took 4 minutes 9 seconds.
+
+The R-CCS Cloud GH200 route was directly validated on `qc-gh200-01` with
+NVHPC 26.3. The exact energy was `-76.24377593489788 Ha`; the internal
+Davidson FOM was `520.734720 s`, the multiply section was `22.730310 s`, and
+the complete Slurm job took 9 minutes 34 seconds.
+
+The R-CCS Cloud FX700 route was directly validated on `fx29` with GCC 8.5,
+MPICH 4.0, 512-bit SVE, and Fujitsu LAPACK. The exact energy was
+`-76.2429584823075 Ha`; the internal Davidson FOM was `747.636165 s`, the
+multiply section was `33.897903 s`, and the complete Slurm job took 13 minutes
+3 seconds.
+
+The Rikyu route was validated with NVHPC 26.3 on project `rkp00012`. All
+three rows converged to `-76.243776776861 Ha`:
+
+| B200 GPUs | Nodes | Internal Davidson FOM (s) | 4-GPU-relative speedup |
+|---:|---:|---:|---:|
+| 4 | 1 | 565.510 | 1.00x |
+| 8 | 2 | 290.222 | 1.95x |
+| 16 | 4 | 152.144 | 3.72x |
+
+A fresh BenchKit validation rebuilt upstream SBD commit
+`9481f290c2f49d4f8e5df9b0c9c87ea0f7937c2c` through `build.sh`, confirmed the
+effective `cc100` rank-distributed/index-reordered/NCCL configuration, and
+submitted the eight-GPU row through `scripts/test_submit.sh`. The job converged
+to `-76.2437767768609 Ha` with a `290.281226 s` internal Davidson FOM, and
+`scripts/result.sh` produced valid result JSON with source provenance and the
+`mult` timing section. Only this eight-GPU row has been independently re-run
+through `test_submit.sh`; the four- and sixteen-GPU rows above have not.
diff --git a/programs/sbd/build.sh b/programs/sbd/build.sh
new file mode 100755
index 0000000..0b8da9f
--- /dev/null
+++ b/programs/sbd/build.sh
@@ -0,0 +1,102 @@
+#!/bin/bash
+set -euo pipefail
+
+system="$1"
+source scripts/bk_functions.sh
+
+REPO_URL="https://github.com/r-ccs-cms/sbd.git"
+REPO_DIR="sbd"
+BUILD_DIR=""
+ARTIFACT_DIR="${PWD}/artifacts"
+
+mkdir -p "${ARTIFACT_DIR}"
+bk_fetch_source "${REPO_URL}" "${REPO_DIR}" "main"
+cd "${REPO_DIR}"
+
+case "${system}" in
+  RIKYU)
+    module purge
+    module load nvhpc/26.3
+    BUILD_DIR="build-rikyu-nvhpc-thrust-rankdist-nccl"
+    nccl_root="/shared/software/hpc_sdk/Linux_aarch64/26.3/comm_libs/13.1/nccl"
+    cmake -S . -B "${BUILD_DIR}" \
+      -DCMAKE_BUILD_TYPE=Release \
+      -DSBD_GPU_BACKEND=thrust \
+      -DSBD_GPU_ARCH=cc100 \
+      -DSBD_THRUST_SAFE_MPI_ALLREDUCE=ON \
+      -DSBD_USE_RANK_DISTRIBUTION=ON \
+      -DSBD_USE_BLOCK_RANK_DISTRIBUTION=ON \
+      -DSBD_REORDER_INDEX_ARRAY=ON \
+      -DSBD_USE_NCCL=ON \
+      -DCMAKE_CXX_FLAGS="-I${nccl_root}/include" \
+      -DCMAKE_EXE_LINKER_FLAGS="-L${nccl_root}/lib -lnccl"
+    cmake --build "${BUILD_DIR}" --parallel
+    ;;
+  RC_DGXSP)
+    source /etc/profile.d/modules.sh
+    module purge
+    module load system/ng-dgx nvhpc-hpcx-cuda13/26.3
+    BUILD_DIR="build-dgxsp-nvhpc-thrust-rankdist"
+    cmake -S . -B "${BUILD_DIR}" \
+      -DCMAKE_BUILD_TYPE=Release \
+      -DSBD_GPU_BACKEND=thrust \
+      -DSBD_GPU_ARCH=cc120 \
+      -DSBD_THRUST_SAFE_MPI_ALLREDUCE=ON \
+      -DSBD_USE_RANK_DISTRIBUTION=ON \
+      -DSBD_USE_BLOCK_RANK_DISTRIBUTION=ON \
+      -DSBD_REORDER_INDEX_ARRAY=ON \
+      -DSBD_USE_NCCL=OFF
+    cmake --build "${BUILD_DIR}" --parallel
+    ;;
+  RC_GH200)
+    module purge
+    module load system/qc-gh200 nvhpc-hpcx-cuda13/26.3
+    BUILD_DIR="build-gh200-nvhpc-thrust-rankdist"
+    cmake -S . -B "${BUILD_DIR}" \
+      -DCMAKE_BUILD_TYPE=Release \
+      -DSBD_GPU_BACKEND=thrust \
+      -DSBD_GPU_ARCH=cc90 \
+      -DSBD_THRUST_SAFE_MPI_ALLREDUCE=ON \
+      -DSBD_USE_RANK_DISTRIBUTION=ON \
+      -DSBD_USE_BLOCK_RANK_DISTRIBUTION=ON \
+      -DSBD_REORDER_INDEX_ARRAY=ON \
+      -DSBD_USE_NCCL=OFF
+    cmake --build "${BUILD_DIR}" --parallel
+    ;;
+  RC_FX700)
+    module purge
+    module load system/fx700 mpi/mpich-aarch64
+    BUILD_DIR="build-fx700-gcc-mpich-rankdist"
+    fjlib="/opt/FJSVstclanga/cp-1.0.30.01/lib64"
+    cmake -S . -B "${BUILD_DIR}" \
+      -DCMAKE_BUILD_TYPE=Release \
+      -DCMAKE_CXX_COMPILER=mpicxx \
+      -DCMAKE_CXX_FLAGS="-march=armv8.2-a+sve -msve-vector-bits=512" \
+      -DBLAS_LIBRARIES="${fjlib}/libfjlapack.so" \
+      -DLAPACK_LIBRARIES="${fjlib}/libfjlapack.so" \
+      -DCMAKE_CXX_STANDARD_LIBRARIES="${fjlib}/libfj90i.so ${fjlib}/libfj90f.so ${fjlib}/libfjsrcinfo.so -lelf" \
+      -DSBD_USE_RANK_DISTRIBUTION=ON \
+      -DSBD_USE_BLOCK_RANK_DISTRIBUTION=ON \
+      -DSBD_REORDER_INDEX_ARRAY=ON
+    cmake --build "${BUILD_DIR}" --parallel
+    ;;
+  *)
+    echo "Unknown system: ${system}" >&2
+    exit 1
+    ;;
+esac
+
+binary="${BUILD_DIR}/apps/chemistry_tpb_selected_basis_diagonalization/diag"
+if [[ ! -x "${binary}" ]]; then
+  echo "SBD executable not found: ${binary}" >&2
+  exit 1
+fi
+cp "${binary}" "${ARTIFACT_DIR}/diag"
+
+# Stage benchmark inputs from the source clone into artifacts so run.sh can
+# retrieve them on the compute node even when the source tree is absent
+# (cross-build mode). See programs/sbd/data/h2o/README.md for provenance.
+for input_file in fcidump.txt h2o-1em4-alpha.txt h2o-1em5-alpha.txt \
+                  h2o-1em6-alpha.txt h2o-1em7-alpha.txt; do
+  cp "data/h2o/${input_file}" "${ARTIFACT_DIR}/${input_file}"
+done
diff --git a/programs/sbd/data/h2o/README.md b/programs/sbd/data/h2o/README.md
new file mode 100644
index 0000000..aba17bd
--- /dev/null
+++ b/programs/sbd/data/h2o/README.md
@@ -0,0 +1,27 @@
+# SBD H₂O benchmark inputs
+
+The SBD H₂O inputs are not committed to the BenchKit repository. `build.sh`
+stages them from the SBD repository's tracked `data/h2o/` directory into
+`artifacts/` during the build, and `run.sh` reads them from there. This file
+records their upstream provenance and checksums.
+
+- Repository: https://github.com/r-ccs-cms/sbd
+- Upstream revision: `02324eee32a49f3203522d230bcbc34ef032a6a6`
+- Files: `fcidump.txt`, `h2o-1em4-alpha.txt`, `h2o-1em5-alpha.txt`,
+  `h2o-1em6-alpha.txt`, `h2o-1em7-alpha.txt`
+
+SHA-256 checksums:
+
+```text
+a3c2302834a33dce7260e8050a3f5180e05dbba1bb748f3e2f6410a7eacbd94d  fcidump.txt
+858c1ef9d430aafe0c325281c7257c2e6cb0e77310a707c1109c55747db02139  h2o-1em4-alpha.txt
+40b29271726daa2be5a5b181535647164e915646710c2641e46dd1ed078a2ee6  h2o-1em5-alpha.txt
+c36920e57507d27dd49d2b729662cbb5a92afce26d61fb0e044e2691599eb804  h2o-1em6-alpha.txt
+1f7af972b56143a9ae6862c02288668656a5bd3fb561337c430e0e478f7cb716  h2o-1em7-alpha.txt
+```
+
+The upstream SBD repository distributes these files under its Apache-2.0
+license (`LICENSE.txt`). `build.sh` copies them from its `bk_fetch_source`
+clone of the repository, so the BenchKit application runs without committing
+the inputs or relying on private project storage. Set `BK_SBD_INPUT_DIR` to
+use an equivalent project-storage copy instead.
diff --git a/programs/sbd/list.csv b/programs/sbd/list.csv
new file mode 100644
index 0000000..781a511
--- /dev/null
+++ b/programs/sbd/list.csv
@@ -0,0 +1,7 @@
+system,enable,nodes,numproc_node,nthreads,elapse
+RIKYU,yes,1,4,32,0:30:00
+RIKYU,yes,2,4,32,0:30:00
+RIKYU,yes,4,4,32,0:30:00
+RC_DGXSP,yes,1,1,1,1:00:00
+RC_GH200,yes,1,1,1,1:00:00
+RC_FX700,yes,1,4,12,1:00:00
diff --git a/programs/sbd/run.sh b/programs/sbd/run.sh
new file mode 100755
index 0000000..f4b4893
--- /dev/null
+++ b/programs/sbd/run.sh
@@ -0,0 +1,144 @@
+#!/bin/bash
+set -euo pipefail
+
+system="$1"
+nodes="$2"
+numproc_node="$3"
+nthreads="$4"
+n_ranks=$((nodes * numproc_node))
+
+source scripts/bk_functions.sh
+
+RESULTS_DIR="${PWD}/results"
+RUN_DIR="${PWD}/sbd_run"
+INPUT_DIR="${BK_SBD_INPUT_DIR:-${PWD}/artifacts}"
+
+mkdir -p "${RESULTS_DIR}"
+: > "${RESULTS_DIR}/result"
+
+case "${system}" in
+  RIKYU)
+    module purge
+    module load nvhpc/26.3
+    determinant_file=h2o-1em7-alpha.txt
+    experiment=H2O-1em7
+    reference_energy=-76.243776776861
+    energy_abs_tolerance=1.0e-12
+    task_comm_size=1
+    adet_comm_size=2
+    bdet_comm_size=2
+    ;;
+  RC_DGXSP)
+    source /etc/profile.d/modules.sh
+    module purge
+    module load system/ng-dgx nvhpc-hpcx-cuda13/26.3
+    determinant_file=h2o-1em5-alpha.txt
+    experiment=H2O-1em5
+    reference_energy=-76.24373504205295
+    energy_abs_tolerance=1.0e-12
+    task_comm_size=1
+    adet_comm_size=1
+    bdet_comm_size=1
+    ;;
+  RC_GH200)
+    module purge
+    module load system/qc-gh200 nvhpc-hpcx-cuda13/26.3
+    determinant_file=h2o-1em6-alpha.txt
+    experiment=H2O-1em6
+    reference_energy=-76.2437759348979
+    energy_abs_tolerance=1.0e-12
+    task_comm_size=1
+    adet_comm_size=1
+    bdet_comm_size=1
+    ;;
+  RC_FX700)
+    module purge
+    module load system/fx700 mpi/mpich-aarch64
+    determinant_file=h2o-1em4-alpha.txt
+    experiment=H2O-1em4
+    reference_energy=-76.2429584823075
+    energy_abs_tolerance=1.0e-12
+    task_comm_size=1
+    adet_comm_size=2
+    bdet_comm_size=2
+    ;;
+  *)
+    echo "Unknown system: ${system}" >&2
+    exit 1
+    ;;
+esac
+
+for input_file in fcidump.txt "${determinant_file}"; do
+  if [[ ! -f "${INPUT_DIR}/${input_file}" ]]; then
+    echo "SBD input not found: ${INPUT_DIR}/${input_file}" >&2
+    exit 1
+  fi
+done
+if [[ ! -x artifacts/diag ]]; then
+  echo "Required artifact not found or not executable: artifacts/diag" >&2
+  exit 1
+fi
+
+rm -rf "${RUN_DIR}"
+mkdir -p "${RUN_DIR}"
+cp artifacts/diag "${RUN_DIR}/diag"
+cp "${INPUT_DIR}/fcidump.txt" "${INPUT_DIR}/${determinant_file}" "${RUN_DIR}/"
+cd "${RUN_DIR}"
+export OMP_NUM_THREADS="${nthreads}"
+
+diag_args=(
+  --fcidump fcidump.txt
+  --adetfile "${determinant_file}"
+  --method 0
+  --block 10
+  --iteration 4
+  --tolerance 1.0e-8
+  --init 0
+  --shuffle 0
+  --carryover_type 0
+  --rdm 0
+  --task_comm_size "${task_comm_size}"
+  --adet_comm_size "${adet_comm_size}"
+  --bdet_comm_size "${bdet_comm_size}"
+)
+
+if [[ "${system}" == "RC_FX700" ]]; then
+  mpirun -np "${n_ranks}" -bind-to numa ./diag "${diag_args[@]}" \
+    > diag.log 2>&1
+else
+  mpirun -np "${n_ranks}" bash -lc \
+    'export CUDA_VISIBLE_DEVICES=$OMPI_COMM_WORLD_LOCAL_RANK; exec "$@"' \
+    bash ./diag "${diag_args[@]}" > diag.log 2>&1
+fi
+
+davidson_time=$(grep -E 'Elapsed time for davidson ' diag.log | tail -n 1 | awk '{print $(NF-1)}')
+energy=$(grep -E '^ Energy = ' diag.log | tail -n 1 | awk '{print $3}')
+mult_time=$(grep -E 'Elapsed time for mult ' diag.log | tail -n 1 | awk '{print $(NF-1)}')
+
+if [[ -z "${davidson_time}" || -z "${energy}" ]]; then
+  echo "SBD completion markers not found" >&2
+  tail -n 80 diag.log >&2
+  exit 1
+fi
+if ! awk -v actual="${energy}" -v reference="${reference_energy}" \
+  -v abs_tolerance="${energy_abs_tolerance}" \
+  'BEGIN {
+    diff = actual - reference
+    if (diff < 0) diff = -diff
+    scale = reference
+    if (scale < 0) scale = -scale
+    exit !(diff <= abs_tolerance + 1.0e-11 * scale)
+  }'; then
+  echo "SBD energy mismatch: got ${energy}, expected ${reference_energy}" >&2
+  exit 1
+fi
+
+cp diag.log "${RESULTS_DIR}/"
+bk_emit_result --fom "${davidson_time}" --fom-unit s \
+  --fom-version "davidson_internal_s" --exp "${experiment}" \
+  --nodes "${nodes}" --numproc-node "${numproc_node}" \
+  --nthreads "${nthreads}" >> "${RESULTS_DIR}/result"
+if [[ -n "${mult_time}" ]]; then
+  bk_emit_section mult "${mult_time}" >> "${RESULTS_DIR}/result"
+fi
+printf 'SBD energy: %s\n' "${energy}" >&2