Update README.md

README.md

@@ -7,8 +7,13 @@ that previously deployed using other framework. After several days trials and er
 
 First, I compared the performance between the default schedule and the schedule generated by ANSOR.
 You can just check `schedule/run.py` for exporting the two versions of runtime and `cpp` folder for 
-all the C++ inference benchmark details. In a nutshell, on my machine(), the default schedule runtime
-is 27.56ms and the optimized schedule runtime is 23.97ms - almost 13% speedup.
+all the C++ inference benchmark details. In a nutshell, on my machine - a 11th Gen Intel(R) Core(TM) i7-11850H @ 2.50GHz CPU, 
+the default schedule runtime for 1000 loops is **5324ms** and the optimized runtime for 1000 loops is **4853ms** - almost 9% speedup.
+
+While the network forward time for 1000 loops using the [MNN](https://github.com/alibaba/MNN) inference engine is **4485ms** -
+7.6% better than the optmized schedule.  I would say the performance of TVM is amazing. It is on pair with high manually optimized 
+neural network engine while saveing AI engineers lots of time on operator tuning. How this speedup can be applied to other new hardware
+still remains to be seen.
 
 ![result.jpg](https://github.com/digital-nomad-cheng/RetinaFace_TVM/blob/main/result.jpg)
 ## Main Contributions