Nvidia’s Blackwell architecture was announced in 2024 and first arrived in AI calculation accelerator cards that are being competed for by the various players in the sector (Microsoft, OpenAI, Tesla, Meta…). Nvidia has just announced consumer graphics cards based on said architecture so without further ado here they are:
- RTX 5090: 21760 CUDA cores, 32 GB of VRAM at 1792 GB/s, 575W, $1999
- RTX 5080: 10752 CUDA cores, 16 GB of VRAM at 960 GB/s, 360W, $999
- RTX 5070 Ti: 8960 CUDA cores, 16 GB of VRAM at 896 GB/s, 300W, $749
- RTX 5070: 6144 CUDA cores, 12 GB of VRAM at 672 GB/s, 250W, $549
The 5090 and 5080 will be released on January 30 while the 5070s will be released in February. All these cards use GDDR7. The Founders Edition (produced directly by Nvidia) uses a new vapor chamber-based design that allows them to use only two slots. They support PCI Express Gen 5 and include a DisplayPort 2.1b port which supports 8K@165Hz.
Let’s go into a little detail. Nvidia announces that in some cases, the 5090 and 5080 will display twice as much FPS as their predecessors. By “certain cases” understand using DLSS 4 which will be exclusive to the RTX 50xx. DLSS 4 will support Multi Frame Generation which will generate up to three images for each image rendered directly by the engine. DLSS 4 will also greatly improve image sharpness by limiting ghosting, improving anti-aliasing and increasing the level of detail. 75 games and applications will support DLSS 4 for the launch of the 50xx. Obviously, Cyberpunk 2077 will be part of the lot and the 5090 will hit 230 FPS in 4K at full speed with the path tracing.
Nvidia also announced lots of new technologies not necessarily exclusive to the 50xx and among them, what Nvidia calls neural shaders. The idea is to train the shader code and game data on a “small” neural network that runs locally during game execution. This allows you to do pretty crazy things like compress textures at will. stolen and use them in rendering without decompressing them, compress the code of the shaders that make up its materials and simplify indirect light calculations by anticipating how the light will bounce. Support for neural shaders will initially be available only via an Nvidia SDK but will eventually arrive directly in DirectX even if it will only work on Nvidia Tensor Cores.
Nvidia also introduced Reflex 2 which drastically reduces latency by modifying rendering based on mouse movements just before sending the image to the screen. There are also plenty of AI-based things that promise us NPCs that are more autonomous, more intelligent and capable of learning. We want to believe them but we have been told the same thing for 30 years. Nvidia also tells us for the nth time that they have crossed the uncanny valley (this time thanks to AI) but looking at the demo we are right in it…
