Rumor: Nvidia GeForce RTX 4090 and 4080 PCB support up to 24GB GDDR6X

German tech website Igor’sLAB has received photos of an RTX 4090 and 4080 reference PCB. It appears that those cards would support up to 24GB of GDDR6X memory. The AD102 GPU and the current GA102 must also be pin compatible.

It usually knowledgable Igor’sLAB has received pictures of a reference pcb for the AD102 gpu from its sources. It should be used in the GeForce RTX 4090 and RTX 4080. It concerns photos of a bare pcb, from which details about the next generation of RTX video cards can be derived.

For example, Igor’sLAB writes that the reference PCB contains as many as 24 vrm phases for the gpu. That would be necessary for the AD102 GPU, which according to previous rumors will have a high TDP of 600W. Nvidia would use three uP9512 controllers for the VRM layout, each of which can generate eight phases. In addition, there are four stages for the video memory.

There are also twelve soldering points for the GDDR6X vram. That memory is made by Micron and is available in modules of 1 or 2GB each. This means that the AD102 can be equipped with a maximum of 24GB of video memory. Lower-ranking RTX 40 models, for example, could come with 8 or 16 GB of memory, writes Igor’sLAB.

The website also writes that the upcoming AD102 GPU and the current GA102 GPU in the RTX 3090 Ti, 3090 and 3080 are mutually pin compatible. If true, then custom graphics card designs for the RTX 3090 Ti could be partially reused for the RTX 4080 or RTX 4090. That would reduce development time and costs for board partners. Igor’sLAB speculates that the RTX 3090 Ti is some sort of trial run for the upcoming AD102 cards with higher 600W tdp. Manufacturers would use a GA102 chip with custom 600W BIOS to test their designs for the upcoming AD102 GPUs, the tech website writes.

Nvidia’s upcoming Lovelace architecture and AD102 consumer GPU is expected in the second half of this year. The cards are expected to have more CUDA cores and L2 cache than their predecessors. Recently, Nvidia already introduced its new Hopper architecture, but it is specifically intended for HPC and data centers and is not coming to consumer video cards.

A diagram of the AD102 reference PCB, based on photos. Source: Igor’sLAB