diff --git a/torchao/sparsity/prototype/superblock/README.md b/torchao/sparsity/prototype/superblock/README.md
index 91110735a..b78bd7875 100644
--- a/torchao/sparsity/prototype/superblock/README.md
+++ b/torchao/sparsity/prototype/superblock/README.md
@@ -45,13 +45,11 @@ At least one GPU:
 
 ## Benchmarking
 
-For all our benchmarking results, you can run `benchmark.sh`. This will run benchmarks with random weights, only testing speedup.
-
+For all our benchmarking results, you can run `benchmark.sh`.
 These benchmarks were run on a NVIDIA-A10080GB, with cuSPARSELt v0.5.2.
 
 
 
-
 ## Training
 Please refer to [TRAINING.md](TRAINING.md) for training from scratch. We use [Torchvision](https://github.com/pytorch/vision/tree/main/references/classification) as our framework for training. Supermask can be applied during training.
 
diff --git a/torchao/sparsity/prototype/superblock/benchmark_results.txt b/torchao/sparsity/prototype/superblock/benchmark_results.txt
index 2a94f8455..5197d0cb1 100644
--- a/torchao/sparsity/prototype/superblock/benchmark_results.txt
+++ b/torchao/sparsity/prototype/superblock/benchmark_results.txt
@@ -1,4 +1,8 @@
-vit_b_16,256,bfloat16,bsr,64,0.8,True,False,52.2149365234375,19.151608075806703
-vit_b_16,256,bfloat16,bsr,64,0.8,True,False,54.3076318359375,18.41361823732223
 model,batch_size,dtype,sparsity,bsr,sparsity_level,quantization,tune_kernel_params,latency,img/s
 vit_h_14,256,bfloat16,None,None,0.0,False,False,489.89296875,2.0412622017245474
+vit_h_14,256,bfloat16,None,None,0.0,True,False,452.5438671875,2.2097305311303126
+vit_h_14,256,bfloat16,semi_structured,None,0.0,False,False,457.6992578125,2.184840772474352
+vit_h_14,256,bfloat16,bsr,64,0.8,False,False,361.4048046875,2.7669803694632154
+vit_h_14,256,bfloat16,bsr,64,0.84,False,False,343.2491015625,2.913336103278675
+vit_h_14,256,bfloat16,bsr,64,0.9,False,False,315.3134375,3.1714474585308468
+vit_h_14,256,bfloat16,semi_structured,None,0.0,True,False,434.479140625,2.301606467370323
diff --git a/torchao/sparsity/prototype/superblock/evaluate.py b/torchao/sparsity/prototype/superblock/evaluate.py
index e10953348..aa181f85c 100644
--- a/torchao/sparsity/prototype/superblock/evaluate.py
+++ b/torchao/sparsity/prototype/superblock/evaluate.py
@@ -59,9 +59,37 @@ def main(args):
     accelerate_with_sparsity(model, args)
 
     criterion = torch.nn.CrossEntropyLoss(label_smoothing=args.label_smoothing)
-    evaluate(model, criterion, data_loader_test, device=device, dtype=torch.bfloat16)
+    return evaluate(model, criterion, data_loader_test, device=device, dtype=torch.bfloat16)
 
 
 if __name__ == "__main__":
     args = get_args_parser(evaluate=True).parse_args()
-    main(args)
+    result = main(args)
+    header = [
+        "model",
+        "batch_size",
+        "dtype",
+        "sparsity",
+        "bsr",
+        "sparsity_level",
+        "quantization",
+        "top-1 acc",
+    ]
+    result_string = ",".join(
+        str(_)
+        for _ in [
+            args.model,
+            args.batch_size,
+            args.dtype,
+            args.sparsity,
+            args.bsr,
+            args.sparsity_linear,
+            args.quantization,
+            result,
+        ]
+    )
+    with open("evaluation_results.txt", "a") as f:
+        if args.header:
+            f.write(",".join(header) + "\n")
+        f.write(result_string + "\n")
+    print(result_string)