Nvidia's new GPUs push performance and price limits

Nvidia’s founder and CEO, Jensen Huang, delivered a keynote speech to kick off the CES 2025 show in Las Vegas. He also unveiled Nvidia's latest RTX 50 GPUs.

Over the past year, the company has frequently received media attention, and its stock price briefly surpassed Apple’s to become the most valuable company in the world.

This growth was primarily driven by its AI chips. However, it was the company’s graphics processing units (GPUs) for gaming and their parallel processing capabilities that laid the foundation for AI advancements.

During the keynote, Huang made several announcements, including unveiling the GeForce RTX 50 GPU family.

“The GPU is just a beast,” he said.

The new line will comprise four graphic cards, RTX 5090, RTX 5080, RTX 5070 Ti, and RTX 5070.

The flagship model, the GeForce RTX 5090, features 92 billion transistors, provides over 3,352 trillion AI operations per second (TOPS) of computing power, 1.8 TB/s bandwidth, and twice outperforms the previous GeForce RTX 4090.

The Blackwell generation of NVIDIA Max-Q technology extends battery life by up to 40%.

Meanwhile, DLSS4 debuts Multi Frame Generation, which boosts frame rates by using AI to generate up to three frames per rendered frame and can increase performance by up to 8x over traditional rendering.

Alongside GeForce RTX 50 Series GPUs, NVIDIA introduced RTX Neural Shaders, which integrate small AI networks into programmable shaders. The company says this unlocks film-quality materials, lighting, and more in real-time games.

The RTX 5090 GPU will cost $2000, compared to $1600 RTX 4090, and will be available starting January.

The GeForce RTX 5070 Ti GPU with 1,406 AI TOPS and GeForce RTX 5070 GPU with 988 AI TOPS will cost $749 and $549, respectively.

Users will be able to buy laptops with new GPUs this March.

In addition to its chips, Nvidia also revealed what it calls Cosmos 'foundation' models that generate photo-realistic video which can be used to train robots and self-driving cars at a much lower cost than using conventional data.

By creating what is known in the tech industry as "synthetic" training data, the models can help robots and cars understand the physical world similar to the way that large language models helped chatbots generate responses in natural language.