diff --git a/hpvm/test/dnn_benchmarks/README.md b/hpvm/test/dnn_benchmarks/README.md
new file mode 100644
index 0000000000000000000000000000000000000000..26450028d2f348adfcab22b7e8eaa032c3f45757
--- /dev/null
+++ b/hpvm/test/dnn_benchmarks/README.md
@@ -0,0 +1,33 @@
+
+
+# Build DNN benchmarks 
+
+Move to your HPVM build directory 
+
+`cd hpvm/hpvm/build`  
+
+## Build All Benchmarks 
+
+`make dnn_benchmarks -j ${NUM_THREADS}`
+
+
+## Build a Specific CNN 
+
+`make ${BENCH} -j ${NUM_THREADS}`
+
+The ${BENCH} name should match the directory name of the respective benchmark under `hpvm-c` directory 
+
+# Run Benchmarks
+
+The benchmarks are built under: `/hpvm/hpvm/build/tools/hpvm/test/dnn_benchmarks/`
+
+Before running benchmarks set up CUDA/CuDNN paths using `hpvm/hpvm/setup_paths.sh` 
+
+Modify the setup script to point to your local installations. 
+
+The accuracy of the benchmark is dumped to a file named `final_accuracy` 
+
+Ensure that this matches the accuracy of corresponding benchmarks here: https://gitlab.engr.illinois.edu/llvm/hpvm/-/tree/approx_hpvm_reorg/hpvm/projects/keras
+
+
+