What does 384-core mean?

it is said that Jetson Xavier NX has a “384-core NVIDIA Volta GPU with 48 tensor cores”. I want to compare it with my own desktop GPU 1060.

My 1060 has 1280 CUDA cores. Should I compare 1280 with 48 Tensor cores or 384-core?

I have no idea what 384-core means and why on the official website I can’t see anything about tensor cores in my 1060 GPU.

Could anybody kindly give me some help?
Thanks!

I’m the wrong guy to answer this, but I can probably get things started…

CUDA cores have been around for a long time. Think of these as individual compute units capable of lots of threads, and not specialized for matrices.

Tensor cores are newer (starting with Volta?), and work on 4x4 floating point matrices. Anything you would do with a tensor might be accelerated by this. Tensor manipulations are used extensively when making geometry or physics transformations (much of general relativity is described via tensors, and mechanical/civil engineers performing a 3D simulation of how a building or material might behave under load will use tensor fields). If your GPU is not Volta or newer, then it won’t have tensor cores.

2 Likes

Xavier NX has 6,6 Tflops (FP16) and 21 TOPS (INT8). The 1060 is specified with 4,4 TFlops (did not find the precision).

Tensor cores allow - tada - highly efficient tensor operations. The 1060 Pascal architecture does not have tensor cores.

2 Likes