AMD CEO Lisa Su says Moore's Law isn't dead, but has slowed down

[Daniel Sims]

Forward-looking: As the pace of transistor advancements and costs changes, tech companies continue to debate how much time Moore's Law has left. The CEO of AMD recently offered fresh comments on the company's efforts to ensure its products keep seeing improvements from generation to generation.

In a recent interview with Barrons, AMD CEO Lisa Su reiterated the company's belief that Moore's Law isn't dead, only that its pacing has changed. AMD is turning to new techniques and technologies to create performance improvements in upcoming generations of its chips.

Su echoes the widespread sentiment that each new generation of transistors is delivering smaller energy efficiency advancements than before and that density is providing smaller improvements. However, the tech industry's processing needs haven't stopped growing, so new techniques like chiplets and 3D packaging are important.

The comments, once again, refute those of chief competitor Nvidia. Green Team head Jensen Huang has declared the death of Moore's Law multiple times over the last several years. He did so most recently to defend the unusually high prices of the GeForce RTX 40 series graphics cards at launch.

In addition to their rising costs, Huang claims that recent transistor nodes' performance and efficiency improvements have slowed to the point where significant gains in products like GPUs are only attainable by sending more watts into more silicon. The comments explain the high launch price and high power requirements of cards like the RTX 4080. AMD also previously expressed fears that GPUs could eat up to 700W by the middle of this decade.

The latest remarks from Su add to what AMD has previously said on Moore's Law. At a summit late last year, chief technology officer Mark Papermaster mentioned things that can help, like accelerators, GPU acceleration, and specialized functions. He also confirmed chiplets were a response to recent transistor price increases.

Moore's Law – named for Intel co-founder Gordon Moore, who passed away in March – postulated that compute power lowers in cost at a regular pace. The theory has more or less held firm since Moore authored his paper on the subject in 1965, but he recently doubted how long it could persist. In 2005, Moore admitted that transistors can't keep shrinking forever, questioning the law's survival beyond 2025.

Permalink to story.

https://www.techspot.com/news/98609-amd-ceo-lisa-su-moore-law-isnt-dead.html

 

[Dimitriid]

That's just about the most meaningless answer I could think of: we know transistors are getting smaller, the point of Moore's law (As pointless as it is) is that he was very specific about double the density every two years.

If it's not every two years, it's just not Moore's Law. It's Lisa Su's law at best.

 

[Irata]

That's just about the most meaningless answer I could think of: we know transistors are getting smaller, the point of Moore's law (As pointless as it is) is that he was very specific about double the density every two years.

If it's not every two years, it's just not Moore's Law. It's Lisa Su's law at best.

This is what Wikipedia says:

Moore's law is the observation that the number of transistors in an integrated circuit (IC) doubles about every two years.

Meaning that chips have 2x the number of transistors every two years. If you go about it the old way (monolithic 2D) then certainly this will get increasingly difficult.

If, however, you use new methods (chiplets, 3D stacking…) and define ‚IC‘ as an MCM rather than a monolithic IC then this can continue.

It‘s really about having to find a new way of doing things to keep Moore‘s Law going forward. And I think bei g an engineer, Lisa Su understands this.

 

[kira setsu]

This'll sound dumb but.

why not just make bigger chips? the 11900 in my system, physically seems small, so why not just make them a bit bigger and keep up the transistor count race?

 

[neeyik]

why not just make bigger chips? the 11900 in my system, physically seems small, so why not just make them a bit bigger and keep up the transistor count race?

Throwing more transistors into the pot to get larger chips results in each wafer fielding fewer of them -- thus the price has to increase. Not only that, the larger the chip, the worse internal latencies become; the more transistor one uses, the greater the power consumption, and the lower the overall maximum clock ceiling.

Just compare AMD's Navi 23 to its Navi 21 -- both are fabricated on the same node, but the former is 237 mm2 in area, with 11.1 billion transistors, whereas the latter is 520 mm2 in size, with 26.8 billion transistors. The fastest card using the Navi 23 has a maximum clock of 2.64 GHz and 176 W whereas the same for the Navi 21 is 2.31 GHz and 335 W.

Outside of GPUs, there isn't a transistor race anymore, as Intel doesn't state component count for its CPUs.

 

[Gezzer]

That's just about the most meaningless answer I could think of: we know transistors are getting smaller, the point of Moore's law (As pointless as it is) is that he was very specific about double the density every two years.

If it's not every two years, it's just not Moore's Law. It's Lisa Su's law at best.

So the article's assertion that the law is

"postulated that compute power lowers in cost at a regular pace"

Is at best a misstatement and/or misinterpretation, if not outright false.

So it's pretty obvious that Moore's law was dead a while ago. The bigger question is how much longer can fabricators wring more performance out their silicon products? We're starting to bump up to the limits of physics here and even with all these refinements we'll eventually reach a point where it won't be economically viable, if even possible. 2025? Yeah, could be, or 2030, 2035. The date is unimportant, it's the fact that it will happen eventually if we continue down the same path.

 

[takemaru]

That's just about the most meaningless answer I could think of: we know transistors are getting smaller, the point of Moore's law (As pointless as it is) is that he was very specific about double the density every two years.

If it's not every two years, it's just not Moore's Law. It's Lisa Su's law at best.

The transistors aren't getting much smaller anymore. They stopped shrinking around 28 nm, the nanometeres on the newer nodes are more of a "performance"/marketing numbers.

 

[Endymio]

If it's not every two years, it's just not Moore's Law. It's Lisa Su's law at best.

Moore's original law was once per year. Intel, recognizing its marketing value, changed it to once every 18 months, then again to once per two years.