Readers who are unaware of the capabilities of NVLink from NVIDIA, the technology was created in 2014 to be “the world’s first high-speed interconnect,” which would allow for much faster GPU data transfers from the CPU, as high as five to twelve times more rapid than PCIe interconnects (by 2014 standards). NVLink was first used in Pascal architecture and has since been used in Volta, Ampere, and Hopper architectures. Hopper currently uses the NVLink 4.0 interconnect, offering 50Gbit per second transfer rates. NVIDIA has slowly been pushing multi-GPU functionality including SLI & NVLINK away from its consumer parts. The previous generation of RTX 30 GPUs only had NVLINK support on the 3090 Ti & 3090 but the Ada flagship, GeForce RTX 4090, doesn’t feature any NVLINK connection at all. Now there are several reasons why this decision could have been made with one being the general lack of multi-GPU support in current APIs and games and the other being to restrict users to use professional products for content creation and AI tasks as the Ada Lovelace GPUs offer huge amounts of AI and FP32 capabilities.
Hopper GPUs, or NVIDIA H100 graphics cards, are designed for AI-centric data centers. NVLink offers multi-GPU support, which would make sense in situations where researchers would need the highest amount of power from the GPU to assist with machine learning and above-normal workflows to discover advanced scientific breakthroughs and more. In gaming, there is very little need for the strength of multiple graphics cards simultaneously. Also, in 2021, NVIDIA halted the support for SLI drivers to begin support for graphics acceleration natively in all games, such as DirectX 12 or the open-source Vulkan support, starting with the RTX 2000 GPU series. But, for overclocking, multi-GPU support would make more sense. Also, multi-GPU support would be helpful for content creators needing additional rendering power. This change from NVIDIA seems like the company decided before shipping reference boards to partners to make boards that would have the same standards and specifications across models from the RTX 40 series. Website TechPowerUp offered pictures of PCBs from MSI, ASUS, Colorful, Palit, ZOTAC, and Gigabyte. VideoCardz stacked the images on top of each other to display where the parent company of AORUS, Gigabyte, had the connection for the NVLink. No official Gigabyte or NVIDIA statement about why the NVLink was removed before selling the newest GPUs is present. We may never know the reason, but it is interesting nonetheless. It is very likely that Gigabyte could be using the same PCB boards as the RTX 3090 Ti to make their RTX 4090 cards or the second case would be that NVIDIA and AIBs had planned to offer NVLINK support on higher-end cards but the decision was scrapped at the very end. News Sources: Der8auer on YouTube, VideoCardz, NVIDIA Customer Help blog, Gigabyte
