By: coppice (coppice.delete@this.dis.org), August 4, 2018 4:34 am
Room: Moderated Discussions
Michael S (already5chosen.delete@this.yahoo.com) on August 25, 2015 2:58 pm wrote:
> And here you understood why AVX512 is a huge mistake. Core-based
> GPGPU killer absolutely needs AVX-1024. Or wider.
If AVX512 had been available in a wide range of devices at the sort of time we originally expected it would probably have been a reasonable stepping stone to a later AVX1024. The real problem for Intel is it has taken too long for AVX512 to get into a large number of hands, and for a good number of people to be working with it. Its already looking a bit sad and dated, and AVX1024 or more is needed today.
The long term issue for Intel is that from the MPP in the 1960s until today, SPMD is the only approach to HPC which has both felt reasonably nature and easy to work with for programmers, and achieved a high hardware utilisation over a wide range of problems. Will it ever be possible for an enhanced CPU to match what well tuned SPMD cores in a GPU can do with SPMD code.
> And here you understood why AVX512 is a huge mistake. Core-based
> GPGPU killer absolutely needs AVX-1024. Or wider.
If AVX512 had been available in a wide range of devices at the sort of time we originally expected it would probably have been a reasonable stepping stone to a later AVX1024. The real problem for Intel is it has taken too long for AVX512 to get into a large number of hands, and for a good number of people to be working with it. Its already looking a bit sad and dated, and AVX1024 or more is needed today.
The long term issue for Intel is that from the MPP in the 1960s until today, SPMD is the only approach to HPC which has both felt reasonably nature and easy to work with for programmers, and achieved a high hardware utilisation over a wide range of problems. Will it ever be possible for an enhanced CPU to match what well tuned SPMD cores in a GPU can do with SPMD code.