@hwlau There's two things going on. The first, on the GPU they use computation to hide communication -- that's the two "streams" concept they show. And their results stop before it gets to the point that the problem size on each GPU is small enough that it no longer hides the communication
Whereas with the Phi, they aren't overlapping things
The second thing that is unclear is how they do their timing. They report the time it takes to update a single spin in nanoseconds. But did they run for 20 minutes and divide by the number of spin updates? Or did they run for half a second? Or for a single spin update (unlikely)?
So it's difficult/impossible to assess things like how much "start-up" time is part of the problem.
Before marking as duplicate, please read.
I've set up a chat room, and I cannot find the link/button or whatever to invite a specific user to join. A search of meta, and the help system suggests this is possible. Instructions say variously, "from the chat room itself", or "from the info link"....
@hwlau It doesn't take much to get through peer review sometimes. Other times it takes way too much. It depends on how friendly your reviewers are I'm finding. Or how well you are known in the field.
My personal opinion is I should be able to reproduce it without taking their code or data, based on just what's in the paper. But those reviewers didn't seem to require that (and I want to speak well of them, I have a paper under review at the same journal ;) )
Computer Physics Communications is awesome in that way. They have 2 types of papers, one is a general research paper (like that one) which doesn't require code distribution but the entire focus is on the computational models
And the second type is a code submission -- you write the paper about the model and then you actually have to give them a fully documented copy of the code which is put online
But it does tend to scare away some of the "big" names because they can't release their codes. That's why I submitted the first type of paper, without the code release, because I could never release mine
Which some people/labs do, but not many. And since our codes are all tightly coupled together, and we license our main fluid solver for hundreds of thousands of dollars per license per year, we wouldn't get an exception
I'm just wonder if a tag for mesoscopic physics should be welcome ? I guess so :-)
Enjoy up/down-voting for this forgotten topic :-)
Some examples found those days:
All Andreev phenomena
Regimes of Josephson junction
and all the questions about Josephson junctions actually are questions a...
Before marking as duplicate, please read.
I've set up a chat room, and I cannot find the link/button or whatever to invite a specific user to join. A search of meta, and the help system suggests this is possible. Instructions say variously, "from the chat room itself", or "from the info link"....
I'm just wonder if a tag for mesoscopic physics should be welcome ? I guess so :-)
Enjoy up/down-voting for this forgotten topic :-)
Some examples found those days:
All Andreev phenomena
Regimes of Josephson junction
and all the questions about Josephson junctions actually are questions a...
