Do we need tensorcores to have mixed-precision training?
Is it bound to Volta and Turing architectures? Or can it be done in Pascal. I saw the webinar and it only mentions Volta and Turing and wanted to make sure.
Do we need tensorcores to have mixed-precision training?
Is it bound to Volta and Turing architectures? Or can it be done in Pascal. I saw the webinar and it only mentions Volta and Turing and wanted to make sure.
It is actually supported in pascal. I use it daily with my two 1080ti. It will give you a very modest speedup (~10%), but your memory will be doubled.