Nvidia's new GPU is so chonk its interconnect is almost the size of an RTX 3090 chip
Nvidia's new Hopper server GPU architecture has spawned a mammoth chip. And I'm not even talking about the 80 billion transistor H100 graphics processor itself. Nope, the new NVSwitch baked into the DGX 100 server system is reportedly home to some 25 billion transistors—that's just shy of the number inside the GA102 GPU inside the RTX 3090.
That 3rd gen NVLink switch chip of the Hopper generation is a monster of its own. It’s also manufactured at TSMC 4N and has 25 billion transistors. The switch offers 64 NVLink ports.The predecessor had 2 billion transistors, was manufactured in TSMC 12 nm and offered 18 ports. pic.twitter.com/37tG229HGsMarch 25, 2022
The NVSwitch chip is the interconnect which allows multiple instances of NVLink—the direct chip-to-chip connection—to talk and enable lots of Nvidia graphics cards to work together. That means they can pool their resources to power through complex calculations faster than a single GPU, or even two GPUs, would be able to on their own.
Think SLI, but connecting lots of chips and, y'know, effective.
This 3rd Gen version is built on the same TSMC 4N node as the H100 chip itself, and is able to directly connect up to eight GPUs with chip-to-chip bandwidth of 900GB/s and a total aggregate bandwidth of 7.2TB/s.
I'm only really referencing those numbers because they are so damned big, not because they have any bearing on us or PC gaming. The Hopper architecture is one designed purely for the server environment, and the NVSwitch interconnect doubly so.
Discrete GPUs didn't really game well together when SLI and CrossFire were actually a thing, and it's only in the more compute-centric environment that genuine multi-GPU processing exists. But even there you need seriously
Read more on pcgamer.com
