Re: [Kaldi-developers] nnet-train-mpe-sequential virtual memory use

[Kirill K. <kir...@sm...>]

> -----Original Message-----
> From: Tony Robinson [mailto:to...@sp...]
> Sent: 2015-05-06 0001
> 
> "its not a bug, it's feature" of CUDA 6 called unified memory -
> http://devblogs.nvidia.com/parallelforall/unified-memory-in-cuda-6/
> (the idea is automatic transfer from CPU to GPU memory - personally I
> don't like it).

Would you mind sharing why not? I read the link and the UM seems a good feature to me--if used carefully. I got an impression that the synchronization of accesses from GPU and CPU will be a funky thing to implement correctly. Is this what you are referring to, or am I missing other points?

 -kkm