Nvidia’s new GPU is so large that its connection is virtually the same size as the RTX 3090 chip it is based on. According to reports, the NVSwitch chip, which joins a large number of NVLink GPU-to-GPU connections, includes approximately 25 billion transistors.
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.
— Andreas Schilling 🇺🇦 (@aschilling) March 25, 2022
The predecessor had 2 billion transistors, was manufactured in TSMC 12 nm and offered 18 ports. pic.twitter.com/37tG229HGs
Nvidia new Hopper server GPU architecture has resulted in the development of a massive chip. And I’m not even referring to the H100 graphics processor, which contains 80 billion transistors. Nope, the new NVSwitch integrated into the DGX 100 server system is supposedly home to approximately 25 billion transistors, which is just a few billion transistors short of the number of transistors found in the GA102 GPU found in the RTX 3090.
The NVSwitch chip serves as an interconnect, allowing several instances of NVLink, which is a direct chip-to-chip link, to communicate with one another and allow a large number of Nvidia graphics cards to work together. In other words, they can pool their resources to do complex calculations much more quickly than would be possible with a single GPU or even two GPUs working independently.
Consider SLI, but with a large number of chips connected and, y’know, more effective. As with the H100 chip itself, this 3rd Generation version is manufactured on the same TSMC 4N node as the H100 chip itself, and it is capable of directly connecting up to eight GPUs at 900GB/s chip-to-chip, and a total aggregate bandwidth of 7.2TB/s.
These figures are being brought up primarily because they are so staggeringly large, not because they have anything to do with us or PC gaming in general. The Hopper architecture, as well as the NVSwitch connection, were both created specifically for use in a server-centric environment.
When SLI and CrossFire were still a thing, discrete GPUs didn’t really play well together in games, and it’s only in the more compute-centric environment that true multi-GPU processing is possible. In order to get the most out of these high-speed processors, you’ll need very fast connectivity, which is likely why Nvidia crammed the NVSwitch with so many damned 4nm transistors into it in the first place.
For comparison, the GA102 GPU, the most powerful processor in the GeForce family, has 28.3 billion transistors, which is a few billion more than the NVSwitch. Compared to the Navi 21 GPU found in the RX 6900 XT, RX 6800 XT, and RX 6800, the Navi 21 GPU in the RX 6900 XT is slightly closer at 26.8 billion. However, it is still significantly greater than the 17.4 billion transistors of the GA104 processor, which drives the RTX 3070 Ti and some RTX 3060 cards, which is used in the RTX 3070.
It is possible that more powerful processors may be integrated into future graphics cards, and the red team is also planning to bring back the concept of multi-GPU gaming. Not necessarily in the same way as CrossFire, but rather in the same way as AMD’s Ryzen CPUs have a chiplet design, and you’ll be on the correct track.
The key to multi-GPU gaming is for the operating system to recognize all of the GPUs as a single chip, and it appears that AMD has developed an internal GPU interface that will do this and allow its top Radeon RX 7000-series cards to feature a multi-chip architecture. And possibly outmuscle the predicted monolithic design of Nvidia’s forthcoming Lovelace-based RTX 4000-series graphics cards in terms of raw power.
