4 ms·
In a sense, these are essentially arrays of nodes, they just happen to be Tesla GPUs (which you can think of as video cards that don't have a video out). The ad
by hvs 12y ago
In a sense, these are essentially arrays of nodes, they just happen to be Tesla GPUs (which you can think of as video cards that don't have a video out). The advantage of these over individual machines is that the I/O speed between machines is much lower than between CPU/GPU.
- Jtsummers 12y agoFrom 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