The node is shipped with Hou 19 but hidden by default (and thus not officially supported)
You can find it inside the Vellum Solver DOP.

The node has a toggle: Leave on GPU, which when checked "Fill the neighbors array, but do not transfer the values back to the CPU. Enabling this option means CPU code (e.g. VEX) and the Geometry Spreadsheet will not see the values, but OpenCL code can access the array without the cost of copying large neighbor arrays back to the CPU."

Question:
Is there a way to access the array stored in the GPU from a following downstream OpenCL SOP node?
I guess no, once the kernel is done, all the local variables are being purged like in vex, right?

When I tried it with an heavy geo and unchecked "Leave on GPU", Houdini just crashed trying to write down a so much data to attributes.



I attach an example file with the node, so that you don't have to dive inside vellum to grab it.

After playing around with it for a bit I think the leave on GPU option only works in dop context and not in sops.
You can take a look at the example file where the neighbor attribute is correctly accessible in dop context even though it does not have any values on the geometry spreadsheet.