I’m using my Orin as part of an edge cluster and was thinking about getting another one to support simultaneous inferencing of large models. Ethernet is good enough for most things I need, but with two dedicated AI nodes, it would be nice to be able to mesh them together into a proper network fabric, especially as that would enable inferencing even larger models. I saw that RoCE is not supported by the Orin, is Infiniband?
The Bluefield DPU system seems like it would be the perfect solution in theory. Are there any network fabric systems that would be supported?
Hi,
This proposal is good and we will check this. Currently we don’t have plan for it in near feature. Will check if we can include it into long term plan.
No worries! Normally I’d think there’s no chance, but after a bit more research, I think NVIDIA is interested in something similar. I want to make my Orin (both, once I get the other) into a set of Converged Accelerators on the A2 scale, I think. I just saw a talk on that and it’s 100% what I’m after.