hankyoreh

Nvidia founder Jensen Huang

For the past half-century, Moore’s Law has been used to explain the rate of development in the global semiconductor industry. But with this law proving limited when it comes to fine process technology, “Huang’s Law” is emerging as a way of gauging the level of performance development in chips, using AI capabilities as a yardstick. Huang’s Law is also drawing attention for being proclaimed by Nvidia, an emerging heavyweight in the world semiconductor industry. It’s not to be confused with “Hwang’s Law,” an improved version of Moore’s Law proclaimed by then Samsung Electronics President Hwang Chang-gyu around a decade ago.

According to a Wall Street Journal piece on Sept. 19, the semiconductor development roadmap shared by Nvidia founder Jensen Huang is “the new Moore’s Law,” taking the place of the law proclaimed by Intel over a half-century earlier. Huang’s Law basically holds that chip performance is dictated not by semiconductor integration level but by AI processing capabilities, with the speed of performance advancements doubling with each year. Indeed, Nvidia’s chip performance has doubled on average every year between November 2012 and May 2020. Moore’s Law, which was proclaimed in 1965 by Intel co-founder Gordon Moore, states the level of transistor integration on a semiconductor chip doubles roughly every 24 months. Indeed, over the 60 years or so since then, improvements in global semiconductor chip integration have more or less conformed to his predictions.

Changes in Intel, Nvidia, AMD market caps

One of the main reasons Huang’s Law is drawing so much attention is that Moore’s Law has reached a limit in terms of being able to fully explain the rate at which the semiconductor industry has been developing recently. As the fine processing technology to increase the number of transistors per chip has reached the ultra-fine level of 2 to 3 nanometers (where 1 nanometer is one-billionth of a meter), additional improvements have become physically impossible. In February 2016, the British journal Nature reported that the semiconductor industry was poised to officially abandon Moore’s Law and present a new roadmap. In a keynote speech for the CES trade show in January of last year, Jensen Huang declared that Moore’s Law was “finished.”

The second reason has to do with a practical reality where advancements in AI have more of an impact on chip performance than integration level. The graphics processing unit (GPUs) that have been Nvidia’s focus have been used primarily for gaming-oriented high-performance graphics, being well suited to parallel processing duties that require different things to be processed simultaneously, as with image pixel generation. Uses for Nvidia’s outstanding parallel processing capabilities have begun to proliferate as they have drawn notice in areas such information processing for self-driving automobiles and AI image recognition. In the environments of AI and the Internet of Things (IoT), where countless calculations have to be performed smoothly and instantaneously, parallel processing capabilities have become increasingly important. One of the reasons Nvidia decided recently to acquire the UK IT company Arm is because of the anticipated synergy between AI and IoT.

The market is already proclaiming Huang’s Law a winner by decision. On July 8, Nvidia surpassed Intel on the New York stock market for market capitalization, making it the US’ highest-valued semiconductor company. This is a reflection of the contrasting fortunes of the two companies, with Nvidia sustaining strong performance amid an explosion in server demand due to the COVID-19 pandemic, while Intel has been stumbling with mass production of 7-nanometer next-generation chips. Some industry observers are even predicting that Nvidia rather than Intel should be looked at as the gauge for future developments in semiconductors.

The battle is obviously not over yet. In contrast with Moore’s Law and its standard of chip integration level, the AI processing capabilities cited in Huang’s Law do not provide a clear standard. While it is true that development in AI parallel processing capabilities have grown in importance, chip performance still hinges on central processing unit (CPU) capabilities, which are determined in turn by chip integration. Nvidia Senior Vice President of Research Bill Dally himself acknowledged that the results of parallel processing could not be adequately assimilated without improvements to CPUs.

By Koo Bon-kwon, senior staff writer

Please direct comments or questions to [english@hani.co.kr]