Vibeleaderboard
← All Intel
Intel / video

Nvidia CTO Michael Kagan: Scaling Beyond Moore's Law to Million-GPU Clusters

Source
youtube.com
Author
Sequoia Capital
Date

Context

Michael Kagan, chief technology officer of NVIDIA and co-founder of Mellanox, the networking company NVIDIA bought in 2019, argues in a Sequoia Capital interview that the network, not only the chip, sets how well a large GPU cluster performs. To speed up a job you split it across many GPUs, which must exchange partial results; if communication is slow or its delay varies, every GPU waits. Kagan says raw link speed is similar across network technologies, and a narrow spread in delay is what lets a job be split across a thousand GPUs instead of ten.

Kagan describes two layers. Scaling up joins GPUs into one unit that software sees as a single GPU; Kagan says Mellanox technology extended this past one server to a 72-GPU system. Scaling out connects many such units. Kagan also credits Mellanox's BlueField data processing unit, which runs a data center's infrastructure software on hardware separate from customer applications, with shrinking exposure to side-channel attacks. This is an NVIDIA executive's account.

Read the source www.youtube.com
More from Sequoia Capital
Recommended reads
Comments

Checking sign-in…

Loading comments…