Nvidia’s GeForce RTX 4090 graphics card may not support NVLink, but traces can be seen on Gigabyte’s Aorus Master GeForce RTX 4090 board.
legendary overclocker Roman ‘der8auer’ Hartung This week, when I took apart Gigabyte’s Aorus Master GeForce RTX 4090 graphics card (one of the best graphics cards out there), I discovered that the printed circuit board (PCB) has NVLink contact locations. There was no real copper on the grid, but the fact that the company left such a mark was interesting.
In fact, Nvidia has officially stated that their AD102 graphics processor does not support NVLink interconnect. This is because we decided to use the transistor budget and die space for other purposes. Therefore, NVLink is gone from Ada Lovelace client GPUs. No other GeForce RTX 4090 board I’ve seen of him has any trace of NVLink. In fact, our colleague video cards I also checked pictures of RTX 4090 PCBs from Asus, Colorful, MSI, Palit and Zotac (published by: Tech Power Up) No trace of NVLink was found.
There are several possible reasons why Gigabyte left the NVLink pad on its GeForce RTX 4090 graphics card.
The AD102 and GA102 graphics processors are pin-compatible, so Gigabyte could have reused the PCB from the GeForce RTX 3090 Ti for the GeForce RTX 4090 to save development costs (we’ve confirmed this you can’t). I haven’t seen a PCB design for Gigabyte’s Aorus Master GeForce RTX 3090 Ti). The company may have a ton of unused GeForce RTX 3090 Ti PCBs, and scrapping them would be costly, so it makes sense to reuse them.
Alternatively, Nvidia notified Gigabyte too late that the AD102 GPU would not have NVLink, so Gigabyte decided not to repeat the PCB design in order to speed up time to market.
Some may have missed NVLink on the GeForce RTX 4090, but it was actually amazing how this technology even made it to the client GPUs (on some Turing GPUs).
Nvidia’s NVLink interconnect (opens in new tab) Originally high bandwidth (300 GB/s P2P with 6 lanes for Volta, 900 GB/s P2P with 18 lanes for Hopper), low bandwidth between CPU and GPU in systems used for high performance computing (HPC). Designed as a delayed link. ) and data analysis. The technology was built into IBM’s Power 8 processors, but AMD and Intel’s latest CPUs don’t support interconnect, so Nvidia can only use it as a multi-GPU interconnect.
NVLink wasn’t really meant for client devices, but Nvidia built such a link into their client GPUs and used it primarily for professional applications (including graphics and HPC workloads) and for the very wealthy few extreme gamers. Enabled multi-GPU support for
With the AD102 GPU, Nvidia decided that building NVLink into the client GPU made less sense than maximizing performance by increasing the transistor budget for computing. A dual GPU graphics subsystem consisting of two AD102-based RTX 6000 48GB professional graphics cards has benefited HPC and some graphics professionals (think 96GB of graphics memory). Well, Nvidia didn’t think it made enough sense to add NVLinks yet. to AD102.
