3 ms·
From the video in one of the links it seems that it's still nodes containing both CPU (POWER) and GPU (Tesla) (some number of each). But instead of PCI-e, it us
by Jtsummers 12y ago
From the video in one of the links it seems that it's still nodes containing both CPU (POWER) and GPU (Tesla) (some number of each). But instead of PCI-e, it uses a new high-speed bus (NVLink) which allows for direct GPU-GPU communication and for the GPU and CPU to share memory at the same speeds (doesn't rely on PCI-e to connect GPU to CPU).
- pmalynin 12y agoThere is now also an RDMA protocol that understands GPU RAM