Reaction to Packet Pushers AI Discusion
Transcript
10 Improvement on performance if you can make sure all of the gpus are getting that data in a timely fashion and so the re-architecture isn't again about doing AI calculations but it's making sure that all the calculators all the gpus are getting their data in that timely fashion no packet loss you know one of the things Cisco talks about is telemetry assisted ethernet load balancing so they're actually using Telemetry to improve Network performance by making smarter load balancing decisions if we could notify the host or
switches if there's Downstream congestion we could update the forwarding tables to avoid the congestion my point would be is that most AI networks should be built non-blocking so this is a Non-Stop this is a waste of time because you know the only time you're going to have a congestion point is at the server not in the back plane not in the ecmp spine because that should all be non-blocking and I still feel as I said three or four weeks ago when we talk about it is
the dpu has to be the critical component here the only way to make a fabric like this non-blocking and not to have to worry about these things is to actually make sure that the server doesn't send the data so the dpu actually is aware of the actual fabric itself and is not you know throwing data out and going fingers crossed let's hope it gets there because that's not not the way we work in 2023
