diff --git a/userbenchmark/dynamo/dynamobench/common.py b/userbenchmark/dynamo/dynamobench/common.py
index fd3fd7e6c9..20445b3b9d 100644
--- a/userbenchmark/dynamo/dynamobench/common.py
+++ b/userbenchmark/dynamo/dynamobench/common.py
@@ -737,6 +737,11 @@ def maybe_mark_profile(*args, **kwargs):
         row.append(kwargs["compression_ratio"])
         row.append(kwargs["eager_peak_mem"])
         row.append(kwargs["dynamo_peak_mem"])
+
+    if "cache_lookup_latency" in kwargs:
+        headers.append("cache_lookup_latency")
+        row.append(kwargs["cache_lookup_latency"])
+
     if "dynamo_stats" in kwargs:
         for k, v in kwargs["dynamo_stats"].items():
             headers.append(k)
@@ -2678,6 +2683,21 @@ def warmup(fn, model, example_inputs, mode, niters=5):
                     optimized_model_iter_fn, model, example_inputs, "dynamo"
                 )
 
+            if self.args.profile_dynamo_cache_lookup:
+                with torch.profiler.profile(
+                    activities=[torch.profiler.ProfilerActivity.CPU]
+                ) as prof:
+                    with maybe_enable_compiled_autograd(self.args.compiled_autograd):
+                        warmup(optimized_model_iter_fn, model, example_inputs, "dynamo")
+
+                events = list(
+                    filter(
+                        lambda event: "TorchDynamo Cache Lookup" in event.key,
+                        prof.key_averages(),
+                    )
+                )
+                dynamo_cache_lookup_latency = events[0].self_cpu_time_total
+
             compilation_time = dynamo_latency - eager_latency + aot_compilation_time
             compression_ratio = (
                 eager_peak_mem / dynamo_peak_mem if dynamo_peak_mem else 0.0
@@ -2695,6 +2715,10 @@ def warmup(fn, model, example_inputs, mode, niters=5):
                 experiment_kwargs["eager_peak_mem"] = eager_peak_mem
                 experiment_kwargs["dynamo_peak_mem"] = dynamo_peak_mem
                 experiment_kwargs["dynamo_stats"] = dynamo_stats
+                if self.args.profile_dynamo_cache_lookup:
+                    experiment_kwargs[
+                        "cache_lookup_latency"
+                    ] = dynamo_cache_lookup_latency
 
             if experiment.func is coverage_experiment:
                 ok, total = Stats.reset_counters()
@@ -3201,6 +3225,13 @@ def get_example_inputs(self):
         help="Enables compiled autograd on compiled benchmark",
     )
 
+    parser.add_argument(
+        "--profile_dynamo_cache_lookup",
+        "--profile-dynamo-cache-lookup",
+        action="store_true",
+        help="profiles TorchDynamo cache lookup",
+    )
+
     group_fuser = parser.add_mutually_exclusive_group()
     # --nvfuser is now the default, keep the option to not break scripts
     group_fuser.add_argument("--nvfuser", action="store_true", help=argparse.SUPPRESS)