by Micheline Murphy for Cisco Blogs, VIP Perspectives
At this year's Cisco Live in sunny Las Vegas, the only word said more often than "Cisco" was "AI." They even had a brand-new installation in the World of Solutions—the AI Hub. Hidden within the AI Hub behind a truly stunning video wall were several small, intimate spaces to talk to some of the leaders in AI—NVIDIA, AMD, and World Wide Technology. NVIDIA, of course, is currently running the table on AI compute with AMD giving NVIDIA a hard run for its money. World Wide Technology, is Cisco's largest partner, and we have been providing multi-vendor networking expertise on high performance networks for decades. As a network engineer at World Wide Technology, it has been one of my great joys to talk to other engineers about what it takes to support an AI workload.
In this latest installment of …Just for Fun! I explore the particularities of an AI workload and how AI's kinks make special trouble for networks and the engineers who support them.
Is that an Elephant? Non-AI Traffic versus AI Traffic
As an everyday engineer, we have a pretty decent idea of how enterprise traffic behaves.
- It's usually unicast, although there might be some multicast.
- Across an entire enterprise, there is a mix of north-south and east-west traffic, perhaps with the east-west traffic concentrated in a data center.
- Generally, the flows are small, therefore short-lived.
- The traffic is random. That is, there is no synchronicity to the timing of flows.
- Small, short-lived, and asynchronous flows lead to an abundance of entropy in the system.
Sound about right?