Hello, I got the HPL for Fermi architectures, but I was wondering if there is a newer version, for Volta archs? If so, where can I acquire it?
NVIDIA does maintain such a version, but it is not publicly distributed by NVIDIA. It is privately distributed (in binary form only) for certain specific purposes, such as doing a top500 run. If you intend to do a top500 run (e.g. you have a machine that can deliver ~1PF or more from GPU acceleration, currently) then you’re welcome to reach out to the OEM that built the machine for you, or NVIDIA, and we will assist. For example, if this describes your situation, and you have no other contacts, you’re welcome to contact me privately on this forum.
I won’t be able to respond to requests for explanation of this policy, exceptions, etc.
There do exist other GPU HPL implementations available publicly, including the Fermi-era one from NVIDIA you referenced, if you wish to do some searching and experimentation. I won’t be able to assist with those.
1PF or more from GPU acceleration alone excludes too much of the community. I hope those are measured “tensor FLOPs”.
@Robert: Do you have a list of available binaries and corresponding architectures built under this program? I’m looking for binaries for Intel+K40 and AMD+V100 for footing on Top500.
FWIW to the rest of the community: some google-fu found HybridHPL/src/hpl-2.0_FERMI_v15 at master · scamicha/HybridHPL · GitHub , but no other public implementations are obvious.
There aren’t any binaries available publicly.
If you have a system that has the numerical capability to get on the next top500 list based on the GPU complement, please reach out to me privately.