new from nvidia Hopper Server GPU Architecture generated a gigantic chip. And I’m not even talking about the 80 billion transistor H100 graphics processor itself. Nope, the new NVSwitch built into the DGX 100 server system is reportedly home to around 25 billion transistors – that’s just under the number inside the GA102 GPU inside the RTX 3090.
This 3rd generation NVLink switch chip from the Hopper generation is a monster in its own right. It is also manufactured on the 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
O NVS switching chip it’s the interconnect that allows multiple instances of NVLink – the direct chip-to-chip connection – to talk to each other and allow many Nvidia graphics cards to work together. This means they can pool their resources to perform complex calculations faster than a single GPU, or even two GPUs, would be able to do on their own.
Think SLI, but plugging in a lot of chips is, you know, effective.
This 3rd generation version is built on the same TSMC 4N node as the H100 chip itself and is capable of directly connecting up to eight GPUs with 900 GB/s chip-to-chip bandwidth and 7.2 TB/s total aggregate bandwidth. .
I’m actually referencing these numbers because they’re so big, not because they have any bearing on us or PC games. The Hopper architecture is designed exclusively for the server environment, and the NVSwitch doubly interconnects.
Discrete GPUs didn’t work well together when SLI and CrossFire were really one thing, and it’s only in the most compute-centric environment that genuine multi-GPU processing exists. But even there you need high-speed links to get the most out of those high-speed processors, which is probably why Nvidia specified the NVSwitch with so many 4nm transistors.
For reference, the GA102 GPU, the largest chip in the GeForce sphere, comes with 28.3 billion transistors – a few billion more than the NVSwitch. On the AMD side, the Navi 21 GPU running the RX 6900 XT, 6800 XT and RX 6800 is closer to 26.8 billion. But it’s still considerably more than the 17.4 billion transistors on the GA104 chip that powers the RTX 3070 Ti, even some RTX 3060 boards.
So yes, a big old chip.
But we will likely see more big chips coming in the next generation of graphics cards, and so will the red team. bring back the idea of multi-GPU gaming, also. Not necessarily in the same way as CrossFire, though, put more thought into the chiplet design of your Ryzen processors and you’ll be on the right track.
The key to multi-GPU gaming is for the OS to see everything as a single chip, and it looks like AMD has figured out an internal GPU interconnect that will do that and allow its Radeon RX 7000 series flagship cards to sport a multi-chip design. . And it potentially surpasses the monolithic design expected of future Nvidia’s Lovelace-based RTX 4000 series cards.