Advanced search

Message boards : Graphics cards (GPUs) : Using Tensor Cores

Author Message
Profile Chilean
Send message
Joined: 8 Oct 12
Posts: 98
Credit: 385,372,461
RAC: 0
Scientific publications
Message 50636 - Posted: 1 Oct 2018 | 18:13:08 UTC
Last modified: 1 Oct 2018 | 18:19:04 UTC

Since NVIDIA is apparently investing plenty on these new "tensor cores" in their new architectures (Volta and Turing)... is ACEMD planning on leveraging these? IIRC, these Tensor Cores can do half-precision calculations pretty fast.

Send message
Joined: 31 Mar 17
Posts: 1
Credit: 3,732,425
RAC: 0
Scientific publications
Message 50637 - Posted: 1 Oct 2018 | 18:32:07 UTC

Turing is in many ways very interesting. It would be a huge performance leap if we can use the Tensor cores for something useful, but there is more:

First, the Turing SM adds a new independent integer datapath that can execute instructions concurrently with the floating-point math datapath. In previous generations, executing these instructions would have blocked floating-point instructions from issuing.

Simultanious execution of integer and floating-point operations could improve performance a lot. Also, FP16 is now supported at double the performance on all cores, so not only Tensor but also regular CUDA.

GeForce RTX 2080 Ti for example:
14.2 TFLOPS of peak single precision (FP32) performance
28.5 TFLOPS of peak half precision (FP16) performance

Post to thread

Message boards : Graphics cards (GPUs) : Using Tensor Cores