Jensen Huang walked onto the CES 2025 stage in Las Vegas and held up a blacked-out graphics card. “Here it is, our brand-new GeForce RTX 50 series, Blackwell architecture,” the Nvidia CEO told an audience of more than 6,000. “The GPU is just a beast.”
The flagship RTX 5090 carries 92 billion transistors and is rated at 3,352 trillion AI operations per second. Nvidia calls it the most powerful GeForce RTX GPU to date. Huang pointed out that it has two cooling fans, and said “even the mechanical design is a miracle.”
The series succeeds the GeForce 40 line. The chips are manufactured by TSMC on a custom 4N process node and use fourth-generation RT cores for hardware ray tracing and fifth-generation Tensor Cores for AI work.
Release dates
The RTX 5090 and RTX 5080 desktop cards were scheduled to go on sale Jan. 30. The RTX 5070 Ti and RTX 5070 follow in February, and laptop versions are expected in March.
What the AI hardware does
DLSS 4 adds Multi Frame Generation, which Nvidia says can raise performance by up to 8x. Huang said it generates three additional frames for every frame the card calculates. “We’re able to render at incredibly high performance, because AI does a lot less computation,” he said. A second feature, Reflex 2, can cut PC latency by up to 75%.
RTX Neural Shaders use small neural networks to improve textures, materials and lighting during play. RTX Neural Faces and RTX Hair handle real-time face and hair rendering, and RTX Mega Geometry raises the number of ray-traced triangles by up to 100x.
The cards also support FP4 computing, which Nvidia says can boost AI inference by up to 2x while using less memory than previous-generation hardware. They run the AI foundation models Nvidia announced for RTX PCs, offered as NIM microservices. “AI PCs are coming to a home near you,” Huang said.
Huang’s keynote ran 90 minutes and covered more than graphics. He also announced the Cosmos platform for robots and self-driving cars, a Toyota partnership using the DRIVE AGX computer, and Project DIGITS, a compact AI supercomputer built on the GB10 Grace Blackwell Superchip that launches in May.
