diff --git a/tools/perf/benchmark.sh b/tools/perf/benchmark.sh
index 4a7cd24487..2770db3999 100644
--- a/tools/perf/benchmark.sh
+++ b/tools/perf/benchmark.sh
@@ -7,8 +7,8 @@ python hub.py
 
 batch_sizes=(1 2 4 8 16 32 64 128 256)
 large_model_batch_sizes=(1 2 4 8 16 32 64)
-backends=("torch" "ts_trt" "dynamo" "torch_compile" "inductor")
-backends_no_torchscript=("torch" "dynamo" "torch_compile" "inductor")
+backends=("torch" "ts_trt" "dynamo" "torch_compile" "inductor" "tensorrt")
+backends_no_torchscript=("torch" "dynamo" "torch_compile" "inductor" "tensorrt")
 
 
 # Benchmark VGG16 model
diff --git a/tools/perf/perf_run.py b/tools/perf/perf_run.py
index 5a91831fe9..c52fb6ba56 100644
--- a/tools/perf/perf_run.py
+++ b/tools/perf/perf_run.py
@@ -15,6 +15,7 @@
 
 # Importing supported Backends
 import torch
+import torch_tensorrt as torchtrt
 from utils import (
     BENCHMARK_MODELS,
     parse_backends,
@@ -23,8 +24,6 @@
     precision_to_dtype,
 )
 
-import torch_tensorrt as torchtrt
-
 WARMUP_ITER = 10
 results = []
 
@@ -294,29 +293,30 @@ def run_tensorrt(
     input_tensors,
     params,
     precision,
-    is_trt_engine=False,
     batch_size=1,
 ):
-    engine = None
-
-    # If the model file is a TensorRT engine then directly deserialize and run inference
-    # else convert the torch module to a TensorRT engine first and then run inference
-    if not is_trt_engine:
-        compile_settings = {
-            "inputs": input_tensors,
-            "enabled_precisions": {precision_to_dtype(precision)},
-            "truncate_long_and_double": params.get("truncate", False),
-        }
-
-        print("Converting method to TensorRT engine...")
-        with torch.no_grad(), torchtrt.logging.errors():
-            model = torchtrt.ts.convert_method_to_trt_engine(
-                model, "forward", **compile_settings
-            )
-
+    # Export an ONNX model and convert to TRT
+    torch.onnx.export(model.eval().cuda(), tuple(input_tensors), "./tmp.onnx")
+    logger = trt.Logger(trt.Logger.WARNING)
+    builder = trt.Builder(logger)
+    network = builder.create_network(
+        1 << int(trt.NetworkDefinitionCreationFlag.EXPLICIT_BATCH)
+    )
+    parser = trt.OnnxParser(network, logger)
+    success = parser.parse_from_file("./tmp.onnx")
+    if not success:
+        raise ValueError("ONNX conversion failed")
+
+    config = builder.create_builder_config()
+    if precision == "fp16":
+        config.set_flag(trt.BuilderFlag.FP16)
+    start_compile = time.time_ns()
+    serialized_engine = builder.build_serialized_network(network, config)
+    end_compile = time.time_ns()
+    compile_time_s = (end_compile - start_compile) / 1e9
     # Deserialize the TensorRT engine
-    with trt.Logger() as logger, trt.Runtime(logger) as runtime:
-        engine = runtime.deserialize_cuda_engine(model)
+    with trt.Runtime(logger) as runtime:
+        engine = runtime.deserialize_cuda_engine(serialized_engine)
 
     print("Running TensorRT for precision: ", precision, " batch_size : ", batch_size)
     iters = params.get("iterations", 20)
@@ -351,7 +351,7 @@ def run_tensorrt(
             meas_time = end_time - start_time
             timings.append(meas_time)
 
-    recordStats("TensorRT", timings, precision, batch_size)
+    recordStats("TensorRT", timings, precision, batch_size, compile_time_s)
 
 
 # Deploys inference run for different backend configurations
@@ -427,11 +427,10 @@ def run(
             )
         elif backend == "tensorrt":
             run_tensorrt(
-                model,
+                model_torch,
                 input_tensors,
                 params,
                 precision,
-                is_trt_engine,
                 batch_size,
             )
         elif backend == "dynamo":
@@ -440,9 +439,6 @@ def run(
         elif backend == "torch_compile":
             run_torch_compile(model_torch, input_tensors, params, precision, batch_size)
 
-        elif backend == "torch_compile":
-            run_torch_compile(model_torch, input_tensors, params, precision, batch_size)
-
         elif backend == "inductor":
             run_inductor(model_torch, input_tensors, params, precision, batch_size)
 
diff --git a/tools/perf/requirements.txt b/tools/perf/requirements.txt
index d204d3c335..159e6f5eab 100644
--- a/tools/perf/requirements.txt
+++ b/tools/perf/requirements.txt
@@ -1,7 +1,9 @@
 numpy
 argparse
 pyyaml
+onnx
 transformers==4.33.2
 diffusers==0.21.4
 pandas==2.0.1
 timm==0.9.8
+