<div dir="ltr"><div dir="ltr"><br></div><br><div class="gmail_quote"><div dir="ltr" class="gmail_attr">On Fri, Feb 21, 2020 at 4:51 PM Junchao Zhang via petsc-dev <<a href="mailto:petsc-dev@mcs.anl.gov">petsc-dev@mcs.anl.gov</a>> wrote:<br></div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex"><div dir="ltr"><div>Hello,</div><div><br></div><div>I want to evaluate MatMult on GPU. I took a 2M x 2M matrix and ran with 6 mpi ranks and 6 GPUs. It took about 0.9 seconds. A kernel launch or a stream synchronization took about 10us. </div></div></blockquote><div><br></div><div>Your call, but you should run the code once and then run it in a new timer. I've seen some big "warm up costs" GPUs today.</div><div> </div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex"><div dir="ltr"><div>Compared with MatMult, they are tiny. Does it mean we can ignore them? What is a proper size to evaluate MatMult? </div></div></blockquote><div><br></div><div>It depends on the purpose/audience for the study. There is no right size other than being much larger than the launch cost, perhaps.</div><div> </div><blockquote class="gmail_quote" style="margin:0px 0px 0px 0.8ex;border-left:1px solid rgb(204,204,204);padding-left:1ex"><div dir="ltr"><div> I heard it is a few thousand rows per MPI rank. Why?</div><div>Thanks.</div><div><div dir="ltr"><div dir="ltr">--Junchao Zhang</div></div></div></div>
</blockquote></div></div>