The upcoming RTX 5070 Ti is built on NVIDIA’s advanced “GB203” silicon, which is the same 4nm chip as the elusive and overpriced RTX 5080. 5070ti however should be available widely, specs wise it includes 8,960 CUDA cores, derived from 70 out of 84 SMs, along with 280 Tensor cores and 70 RT cores. The card is also equipped with 16 GB of GDDR7 memory running at an minimum speed of 28 Gbps. For those wondering about its AI Prowess, it will be delivering a solid 896 GB/s of memory bandwidth, here is memory bandwidth comparison chart of the 5000 series.
We have also provided expected local inference numbers of the 5080 here, you can expect around 15% less performance in tok/s. Since it is also a similar 16GB VRAM part there is absolutely no reason to go with 5080 now, gamers and AI hobbyist alike will be able make use of it particularly well. In case of local inference two of them will provide effective 32GB of VRAM for even higher 32B parameter models and cheaper overall when compared to the RTX 5090.
To run the RTX 5070 Ti, you’ll need to connect it with two PCIe 8-pin power cables, adapter will be inside the box but since there is no founder edition you may very well can plug the 8 pin cable directly depending upon the manufacturer. The graphics card consumes 300 watts of power, so it’s essential to have at least a 750-watt power supply in your system for stable performance.