Jetson AGX Xavier performance. TFLOPS or TMACS?


I want to test whether the deep learning model can use all the gpu's performance. 

From It said that Jetson AGX Xavier have a 11 TFLOPS FP16 gpu. But CNN convolution operation is often evaluated as TMACS.  

So for Jetson AGX Xavier. Its full power should be 5.5 TMACS or 11 TMACS?

Moving this thread to the Xavier forum.

The Jetson AGX Xavier performance is in TFLOPS for FP16 and TOPS for INT8. For more info about Xavier GPU and DLA in FP16/INT8, please refer to this post:

Thanks for moving the question to right forum.

Yes I know performance is in TFLOPS for FP16.

I know TMACS = 5.5T plus + 5.5T mult = 11TFLOPS

And gpu can implement multiply–add instrument. So 5.5TMACS can equal 5.5M multiply–add.

So my question is that:

11 TFLOPS means 5.5M TMACS or 11 TMACS. I assume 5.5M TMACS is answer but I didn’t find official reference about that.

Thanks for any clarification or some references about it.

Hi, believe it is 5.5 since one MAC = two FLOP as you pointed out.