Nvidia's new GPU is so chonk its interconnect is almost the size of an RTX 3090 chip

According to reports, the NVSwitch chip, which joins a large number of NVLink GPU-to-GPU connections, has roughly 25 billion transistors. Golly.

The new Hopper server GPU architecture from Nvidia has resulted in a massive chip. I'm not even talking about the H100 graphics processor's 80 billion transistors. No, the new NVSwitch in the DGX 100 server system is said to have 25 billion transistors, which is just a few billion less than the GA102 GPU in the RTX 3090.

The NVSwitch chip acts as an interconnect, allowing several instances of NVLink—a direct chip-to-chip connection—to communicate and enabling multiple Nvidia graphics cards to work together. That means they can pool their resources to do complicated calculations more quickly than a single GPU, or even two GPUs, could do on their own.

Consider SLI, but with a lot of chips connected and, well, effective.

This 3rd Gen version is based on the same TSMC 4N node as the H100 chip, and it can connect up to eight GPUs directly with 900GB/s chip-to-chip bandwidth and 7.2TB/s total aggregate bandwidth.