Is communication speed between GPUs important?



Nvidia’s DGX-1 is specialized server for deep learning.
Its GPU-GPU speed is high level, but CPU-GPU is normal (Not NVLink).
Using Power, it is having NVLink, so CPU-GPU spped will be high level.
But we think Power is not standard for server currently.
Running MapD system, is it important to communicate with GPUs directly?
If so, we think we can choice DGX-1, not Power.
We’d like to know if it is a good choice to use DGX-1 for MapD.



Hi @kiuchi, thanks for trying out MapD.

Both DGX-1 and Power are great platforms for MapD. You can build the open source code for Power currently, however a supported build of the MapD Community and Enterprise editions is on the near-term roadmap.

We definitely see speedup by leveraging NVLink between host and the GPUs that the Power Minsky server provides. As for the DGX-1, we do not leverage the NVLink between the GPUs yet (except for certain scenerios around rendering) but are considering adding accelerated inter-GPU reduction in the future that would leverage inter-GPU NVLink.


Hi @darwin, thanks for reply.

I’m glad to hear that standard intel motherboard is great platform, too.
And I’ll try to check Power chip in the future.