Huang Renxun: Nvidia surpasses Moore's Law, chip performance increases 1,000 times in 10 years

Source: AI Fan

Nvidia CEO Jensen Huang said the company's AI chip performance has advanced faster than Moore's Law, the maxim that has dominated computing for decades.

"Our systems are developing much faster than Moore's Law," Huang said in an interview the morning after delivering a keynote to 10,000 people at the CES conference in Las Vegas.

Moore's Law, which was proposed by Intel co-founder Gordon Moore in 1965, predicts that the number of transistors on a chip will double about every year, thereby doubling the chip's performance. This prediction has largely been realized in the decades that followed, driving rapid increases in computing power and sharp declines in costs.

While the momentum of Moore's Law has slowed in recent years, Huang said Nvidia's AI chips are developing at a faster pace.The company says its latest data center superchip has more than 30 times the performance of the previous generation in AI inference.

Huang explained that by innovating at all levels simultaneously, including architecture, chips, systems, libraries and algorithms, it is possible to break through the limitations of Moore's Law.

The Nvidia CEO made this bold claim at a time when many people are questioning whether AI development has stagnated. Top AI labs such as Google, OpenAI and Anthropic are using Nvidia's AI chips to train and run their models, and the improvement in chip performance is likely to bring further breakthroughs in AI capabilities.

This is not the first time Huang Renxun has said that Nvidia is beyond Moore's Law. He mentioned in a podcast in November that the field of AI is experiencing a "super Moore's Law" type of development.

He refuted the claim that AI progress is slowing down and pointed out that there are currently three laws of AI development: pre-training (learning patterns from massive amounts of data), post-training (fine-tuning through human feedback and other methods) and test-time computing (giving AI more "thinking" time).

Huang said that just as Moore's Law has driven the development of computing technology by reducing the cost of computing, improving AI inference performance will also reduce its cost of use.

While Nvidia's H100 was once the chip of choice for technology companies to train AI models, as these companies shifted their focus to the inference stage, some began to question whether Nvidia's expensive chips can continue to maintain their advantage.

Currently, AI models that use test-time computing are expensive to run. Take OpenAI's o3 model, for example, which achieves human-level performance in the General Intelligence Test, but costs nearly $20 per task, while the ChatGPT Plus subscription fee is $20 per month.

In his keynote on Monday, Huang showed off the latest data center super chip, the GB200 NVL72. The chip has a 30-40 times improvement in AI inference performance over the previous best-selling H100. He said this performance leap will reduce the cost of using models like OpenAI o3 that require a lot of inference computing.

Huang stressed that their focus is on improving chip performance because in the long run, better performance means lower prices.

Increasing computing power is a direct solution to the performance and cost issues of computing at test time, he said. In the long run, AI inference models can also provide better data for pre-training and post-training stages.

In the past year, thanks to computing breakthroughs from hardware companies such as Nvidia, the price of AI models has indeed dropped significantly. Although OpenAI’s latest inference model is expensive, Huang expects this price reduction trend to continue.

He also said that Nvidia’s AI chip performance today is 1,000 times higher than it was 10 years ago, which is far faster than the development of Moore’s Law, and this rapid development momentum shows no signs of stopping.