How the 32 TOPs of Jetson AGX Xavier is calculated?

[jerryahn]

This site (https://developer.nvidia.com/embedded/develop/hardware) shows that AGX Xavier supports 32 TOPs. I think it’s huge. I wonder how this speed was calculated.
In my calculations, this device has 512 1.3GHz GPU cores, which is equivalent to 666 GOPs. In addition it has 64 1.3GHz Tensor cores, each of which can perform 64 operations simultaneously, resulting in 4.92TOPs. The sum of the two yields 5.58 TOPs. I think I’m missing something. What’s the right calculation?

[dusty_nv]

Hi jerryahn, please refer to this FAQ for Jetson AGX Xavier performance: https://developer.nvidia.com/embedded/faq#xavier-performance

Jetson AGX Xavier is capable of more than 30 TOPS (trillion operations per second) for deep learning and computer vision tasks. The 512-core Volta GPU with support for Tensor Cores and mixed-precision compute is capable of up to 11 TFLOPS FP16 or 22 TOPS INT8 compute. Jetson AGX Xavier’s dual NVDLA engines are capable of 5 TOPS INT8 or 2.5 TFLOPS FP16 performance each. It also has high-performance eight-core ARM64 CPU, a dedicated image processor, a video processor and a vision processor for accelerating computer vision tasks.

[jerryahn]

I think I overlooked precision. I calculated based on single-precision only.
Amazing resource it supports!
Thanks you.