[enriquto]

> I don't see any claims about performance, but I would be very surprised if it was anything better than abysmal. In a modern neural network pipeline, just sending data to the CPU memory

Notice that the bulk of data does not necessarily go through the pipeline (and thus by the cpu). You may only send a "token", than the program downstream uses to connect to and deal with the actual data that never left the gpu.

[ken]

This is similar to the Tcl approach. “Everything is a string”, but the string may be something like “the network socket at address 0x12345678”.

[jasonhansel]

Would be nice if Unixes provided better support for pipelines of this sort...

[bollu]

Pretty sure you can't do this since the GPU's memory is per-process isolated as well.

[tntn]

Unless you share it between processes.