Dell Technologies Inc. (DELL) Discusses Scaling AI Networks From 800G to 1.6T and Key Differences From Traditional Data Centers August 27, 2026 1:30 AM EDT
Company Participants
Linas Dauksa
Avik Bhattacharya
Ryan Harris
James Wynia
Presentation
Linas Dauksa
Welcome to this webinar on 800 gig to 1.6T, Making AI Networks Work at Scale. My name is Linas Dauksa from Keysight Technologies. I’m joined by Jim from Dell; Ryan from Siemon; and Avik from Keysight.
Question-and-Answer Session
Linas Dauksa
Let’s jump into the very first question here. AI infrastructure performance depends on the entire system. What fundamentally changes compared to traditional data center networks? Who’d like to take it?
Avik Bhattacharya
Yes. Maybe I can start with it. So the traditional networking is more like optimizing for flexibility of traffic, more general purpose traffic, whereas the AI networking in today’s world is more for optimizing those GPU to GPU elephant flows, the GPU to GPU training traffic, right? I mean, I can take an example of, let’s say, TCP, right, our good old friend TCP, it backs off when there is congestion, right? It’s good for general purpose traffic, good for web traffic, but maybe catastrophic for AI traffic, right? So that’s why, in AI networking, we use technologies like PFC or DCQCN with RoCEv2 to do flow control, which is — gives more network utilization with respect to GPU-to-GPU training traffic.
Linas Dauksa
And when you say large elephant flows, you mean large flows that are difficult to load balance, correct?
Avik Bhattacharya
That’s exactly right. Large flows created by multiple GPUs talking at the same time, and they’re hitting the switch all at the same time creates a microburst.
Ryan Harris
I was going to add something from the cabling perspective. Traditional
