Gordon Moore, the co-founder of Intel, one of the most prestigious engineers and technological entrepreneurs, passed away a few weeks ago. His observation that the number of transistors in an integrated circuit (IC) doubles about every two years later became known as “Moore’s law”. This law had holden true for nearly half century and had been a major driving-force for the performance scaling of computing systems.
In this blog post, I would like to discuss Moore’s law, the economics of Moore’s law, and why Moore’s law is no longer valid.
The number of transistors in an IC every year corresponds to the green curve in the following figure from AMSL’s Investor Day 2021 Event Presentation. We could see that Moore’s law held perfectly well before 2005. However, since around year 2005, the transistor scaling has slowed down and is predicted to be slowed down in the future. The performance scaling could still be maintained to some extent because of the system scaling technologies, such as advanced packaging, despite the performance scaling brought by transistor scaling has slowed down.
Therefore, the original Moore’s law no longer holds true nowadays. Although NVIDIA CEO Jensen Huang considered Moore’s law dead publicly in 2022, he might not be the first person who was aware of this. At least some other less well-known folks, such as Doug O’Laughlin, had already mentioned Moore’s law was dead in 2020 or even earlier.
The original Moore’s law talks nothing about the cost. Even if Moore’s law can hold true for another several decades, if the cost per transistor in an IC remains the same every two years and the performance boost brought by the system scaling technologies such as advanced packaging is small year over year, the ordinary consumer would not get a performance boost using the same budget every two years, and only the high-end consumer could get a 2x performance boost by paying 2x costs every two years. Certainly, this would be completely not appealing to the majority of the consumers.
From the International Business Strategies analysis presented in the Marvell’s 2020 Investor Day, the cost per transistor used to go down a lot every two years, allowing the customer to pay a little bit more money for a 2x performance boost. However, as the process technology gets to its physical limit, the cost per transistor ceased to go down, primarily due to the cost increase in technological innovations. In fact, the cost per transistor actually increases as the process technology advances. The increase in the cost per transistor canceled lots of the performance boost brought by the system scaling technologies such as advanced packaging. In other words, the performance boost brought by the system scaling technologies such as advanced packaging is not free. This means, even if the original Moore’s law, which was about transistor scaling, can still hold true, it’s no longer meaningful and attractive for ordinary consumers.
From the TSMC wafer prices, we could also infer that the per transistor price ceased to go down. Assuming the transistor density gets increased by 1.5x every two years, the same-size wafer price gets increased by 1.5x every two years as well in recent years, suggesting that the per transistor price remains the same.
Moore’s law was extremely critical for technological advancement and economic growth, which are significantly dependent on the computing capability growth. However, as Moore’s law is dead and its economics is also dead, the consumer could hardly get free performance boost at the same price from transistors process innovations every two years. It’s time to have additional innovations to improve system performance so that the consumer could still get great performance boost at the same price every a few years. Otherwise, the technological and economic growth will also slow down.
NVIDIA believes computing is not only a chip problem but also a software problem. An innovation synergy between hardware, software and artificial intelligence, i.e., the entire system stack, will maintain a new Moore’s law in the future.