mapesdhs
Distinguished
They wouldn't use the Titan X though, because as I said earlier, it does not have ECC and numerous
other features which are essential for the sort of task run at such insitutions. Teslas have a full speed
PCIe return path (gamer cards don't), various caching features not found on gamer cards, etc. It's a
broader issue than just the degree of 64bit fp strength.
Note that only certain types of supercomputer-level compute task are suitable for GPU acceleration,
because of the complexity of inter-node communication. Striking the right fine/course-grain balance
is critical, hence the existence of shared-memory systems like the UV2 which can work much better
when a lot of data has to be managed aswell as the raw compute load, though such systems can
exploit GPUs aswell. I'm still stuck with a lowly 36-CPU Onyx3800 IR4, so if you have a spare Titan
SC then feel free to send it over. ;D
Ian.
PS. Highly likely that behind the scenes NVIDIA makes custom undisclosed GPUs for DARPA, Lockheed
and others anyway. SGI did the same thing 20 years ago, doubt much has changed. Probably expt
with custom multi-GPU arrays, units with a lot more RAM (forget 12GB, I mean more like 300GB), etc.
Most of SGI's gfx people moved to NVIDIA.
other features which are essential for the sort of task run at such insitutions. Teslas have a full speed
PCIe return path (gamer cards don't), various caching features not found on gamer cards, etc. It's a
broader issue than just the degree of 64bit fp strength.
Note that only certain types of supercomputer-level compute task are suitable for GPU acceleration,
because of the complexity of inter-node communication. Striking the right fine/course-grain balance
is critical, hence the existence of shared-memory systems like the UV2 which can work much better
when a lot of data has to be managed aswell as the raw compute load, though such systems can
exploit GPUs aswell. I'm still stuck with a lowly 36-CPU Onyx3800 IR4, so if you have a spare Titan
SC then feel free to send it over. ;D
Ian.
PS. Highly likely that behind the scenes NVIDIA makes custom undisclosed GPUs for DARPA, Lockheed
and others anyway. SGI did the same thing 20 years ago, doubt much has changed. Probably expt
with custom multi-GPU arrays, units with a lot more RAM (forget 12GB, I mean more like 300GB), etc.
Most of SGI's gfx people moved to NVIDIA.