diff --git a/docs/superbench-config.mdx b/docs/superbench-config.mdx
index feebef6d6..2d3454ee6 100644
--- a/docs/superbench-config.mdx
+++ b/docs/superbench-config.mdx
@@ -370,6 +370,7 @@ proc_num: int
 node_num: int
 env: dict
 mca: dict
+bind_to: string
 prefix: str
 parallel: bool
 ```
@@ -403,6 +404,7 @@ Some attributes may only be suitable for specific mode.
 | `prefix`   |    ✓    |          ✘          |   ✘   |
 | `env`      |    ✓    |          ✓          |   ✓   |
 | `mca`      |    ✘    |          ✘          |   ✓   |
+| `bind_to`  |    ✘    |          ✘          |   ✓   |
 | `parallel` |    ✓    |          ✘          |   ✘   |
 | `pattern`  |    ✘    |          ✘          |   ✓   |
 
@@ -452,6 +454,16 @@ MCA (Modular Component Architecture) frameworks, components, or modules to use i
 in a flatten key-value dictionary.
 Only available for `mpi` mode.
 
+### `bind_to`
+
+Process binding policy passed to `mpirun -bind-to`.
+Only available for `mpi` mode.
+
+Use this option when a benchmark needs to override the runner's default MPI binding behavior,
+for example when the benchmark implements its own topology-aware CPU/NUMA affinity logic.
+
+* default value: `numa`
+
 ### `parallel`
 
 Whether run benchmarks in parallel (all ranks at the same time) or in sequence (one rank at a time).
diff --git a/superbench/runner/runner.py b/superbench/runner/runner.py
index a5ac13cbb..777f1b5b3 100644
--- a/superbench/runner/runner.py
+++ b/superbench/runner/runner.py
@@ -91,6 +91,8 @@ def __validate_sb_config(self):    # noqa: C901
                             'btl_tcp_if_exclude': 'lo,docker0',
                             'coll_hcoll_enable': 0,
                         }
+                    if 'bind_to' not in mode:
+                        self._sb_benchmarks[name].modes[idx].bind_to = 'numa'
                     for key in ['PATH', 'LD_LIBRARY_PATH', 'SB_MICRO_PATH', 'SB_WORKSPACE']:
                         self._sb_benchmarks[name].modes[idx].env.setdefault(key, None)
                     if 'pattern' in mode:
@@ -182,13 +184,14 @@ def __get_mode_command(self, benchmark_name, mode, timeout=None):
                 '-tag-output '    # tag mpi output with [jobid,rank]<stdout/stderr> prefix
                 '-allow-run-as-root '    # allow mpirun to run when executed by root user
                 '{host_list} '    # use prepared hostfile or specify nodes and launch {proc_num} processes on each node
-                '-bind-to numa '    # bind processes to numa
+                '-bind-to {bind_to} '    # bind processes according to mode config
                 '{mca_list} {env_list} {command}'
             ).format(
                 trace=trace_command,
                 host_list=f'-host localhost:{mode.proc_num}' if 'node_num' in mode and mode.node_num == 1 else
                 f'-hostfile hostfile -map-by ppr:{mode.proc_num}:node' if 'host_list' not in mode else '-host ' +
                 ','.join(f'{host}:{mode.proc_num}' for host in mode.host_list),
+                bind_to=mode.bind_to,
                 mca_list=' '.join(f'-mca {k} {v}' for k, v in mode.mca.items()),
                 env_list=' '.join(
                     f'-x {k}={str(v).format(proc_rank=mode.proc_rank, proc_num=mode.proc_num)}'