NVIDIA CEO Jensen Huang announced live that the upcoming “Ada” RTX 4090 (24GB G6X) will cost $1,599, the RTX 4080 with 16GB G6X will cost $1,199, and the RTX 4080 with12GB G6X will chime in at $899. Named after Ada Lovelace, the new GPUs feature the 4N process from TSMC, which allowed NVIDIA to fit in 76 billion transistors into the die. With over 18,000 CUDA cores (70% more than Ampere) and clock speeds north of 3GHz achieved in their own labs, we should see up to 25% gaming performance boosts (or up to 4x increases using RTX). Ada also introduces DLSS 3 so that AI can predict an entire upcoming frame instead of just pixels.
Jensen also mentioned that by predicting future frames the CPU will no longer bottleneck the GPU in certain games with a lot of physics calculations such as Microsoft Flight Simulator, which saw a 2x FPS increase with RTX ON versus RTX OFF.
Ada can increase game performances by up to 4x using AI over brute force rendering by the CPU. In DLSS 2, the AI model takes a low resolution current frame and compares it to a 16K resolution previous reference frame to predict, on a pixel by pixel basis, a current high resolution frame. Adding to the DLSS 2 AI model, in DLSS 3 a whole frame can now be predicted by comparing the sequential frames to discover how the scene is changing.
The new Optical Flow Accelerator in Ada gives the neural network pixel vector information to fill in frames and can boost performance up to 4X, and can do this without utilizing the graphic pipeline. We look forward to seeing how the new GPUs perform when they are on shelves and in the hands of the media.
Source: NVIDIA GTC Stream