The modern AI and machine-learning industry has been more or less invented by Nvidia over the past decade. The business continues to make impressive generation-on-generation strides and Ampere’s per dollar output is very strong. Currently, Nvidia does not have any significant competition in the GPU AI market. “But there is no such thing as” Huang’s Rule, “having said all that. That is the appellation given by Christopher Mims of the Wall Street Journal to Nvidia CEO Jensen Huang.
So, what’s the Rule of Huang? “Well, it’s a mistaken description of Moore’s Law, but instead of” Moore “with the word” Huang “in front of it. Specifically: I call it Huang’s Rule, after the chief executive and co-founder of Nvidia Corp., Jensen Huang. It describes how every two years the silicon chips that power artificial intelligence more than double in performance.
Why doesn’t work the Description?
In combining Moore’s Law with Dennard scaling, Mims begins his description. Moore’s Law predicted that every two years, the number of transistors on a chip will double. Dennard Scaling projected that it would reduce their power consumption and allow for faster clocks by building smaller transistors closer together. A calculation of density is Moore’s Law. Dennard Scaling calculates per watt efficiency and applies to absolute performance. It is true that in the colloquial discourse, these two different discoveries are always merged, but in this particular case, combining the two confuses the reality of the situation.
Mims writes: “The rule of Moore slowed down, and some say it’s over.” But a new rule, theoretically no less relevant for the next half-century of computing, has emerged.
The definition of Moore’s Law, as we have mentioned a few times on this page, is nuanced and vulnerable to periodic changes. If you conflate Moore’s Law and Dennard’s scaling mistakenly, Moore’s Law has slowed down a lot. If you consider Moore’s Law strictly as a measure of transistor density, it’s actually kept close to its historical long-term speed. This 1970-2018 map makes it very clear. Dennard Scaling, which ended in about 2004, was what broke down.