Alleged specifications of GeForce RTX 3090 and RTX 3080 are online
The specifications of two Nvidia RTX 3000 cards have been put online by VideoCardz. The site claims the details are from video card manufacturers. The RTX 3090 would get 24GB gddr6x and consume up to 350W.
According to the leaked information from VideoCardz, Nvidia will present three RTX 3000 cards next week: the RTX 3090, 3080 and 3070. The cards will get an Ampere GPU, the second generation of Nvidia’s custom cores for ray tracing, and the third generation of Nvidia’s Tensor- cores.
The RTX 3090, 3080 and 3070 would be equipped with 24GB, 10GB and 8GB vram respectively. A 20GB variant of the RTX 3080 would also be in the works, but that model may not be released until later. The video cards receive support for PCI-E 4.0, HDMI 2.1 and DisplayPort 1.4a.
The top model, the RTX 3090, would be equipped with the GA102-300 GPU with 5248 cores and a 384bit memory bus. In combination with the gddr6x memory, this provides a theoretical bandwidth of 936 GB / s. GeForce RTX 3090 video cards from manufacturers are said to feature two 8-pin connectors and can consume up to 350W. Nvidia itself previously showed a 12-pin connection, which is probably on the Founders Edition.
The RTX 3080 is also based on the GA102 GPU, but a variant with fewer cores. Nvidia would have the GPUs made at 7nm. The RTX 3070 would get a GA104 GPU, but details are not yet known. The new leaked details only concern the RTX 3090 and 3080, the specifications of the RTX 3070 are based on previous rumors.
GeForce RTX 3090 | GeForce RTX 3080 | GeForce RTX 3070 | |
GPU | 7nm, GA102-300 | 7nm, GA102-200 | nm, GA104-300 |
Cuda cores | 5248 | 4352 | nnb |
Boost speed | 1695MHz | 1710MHz | nnb |
Vram | 24GB gddr6x | 10GB gddr6x | 8GB gddr6 |
Memory speed | 19.5GBit / s | 19Gbit / s | 16Gbit / s |
Memory bus | 384bit | 320bit | 256bit |
Memory bandwidth | 936GB / s | 760GB / s | 512GB / s |
Tgp | 350W | 320W | 220W |
Specifications according to VideoCardz, based on custom RTX 3000 cards without overclocking.