Intel CEO Pat Gelsinger says Moore's Law isn't dead, but it has slowed down

[Daniel Sims]

Why it matters: Recent developments in hardware production have intensified the conversation regarding how much time Moore's Law has left. Nvidia has repeatedly declared its death, but AMD and Intel believe it has only slowed down and that numerous innovations can help new products maintain the performance improvements clients expect.

MIT recently posted a video of a talk from earlier this year where Intel CEO Pat Gelsinger commented on recent assertions that Moore's Law could end soon. Gelsinger believes the rule guiding chip production for almost six decades still holds but admits that it hasn't maintained its usual pace lately.

Moore's Law, coined by the late Intel co-founder Gordon Moore in 1965, postulates that the number of transistors per square inch on a circuit board will roughly double every two years. That rule has mostly held firm ever since, enabling reliable performance gains.

Were Moore's Law to end, as Nvidia CEO Jensen Huang has claimed numerous times since 2017, then building faster devices would theoretically require pumping more electricity into more transistors, significantly raising costs and energy consumption. The last time Huang declared time-of-death on Moore's Law in 2022, the assertion was in response to criticism of price increases from the company's RTX 3000 to RTX 4000 graphics cards.

Speaking at MIT, Gelsinger said that the recent rate of transistor doubling has been more like three years instead of two, admitting that the "Golden Age" of Moore's Law is over. The rising costs of semiconductor fabs have been a central factor.

Gelsinger noted that the cost of a modern fab has grown from around $10 billion to $20 billion in the last seven or eight years. Consulting firm IBS recently predicted that a 2nm fab could cost about $28 billion – 50 percent more than a 3nm facility. The rising need for EUV lithography tools is a primary cause behind the growing expenses.

Because performance uplifts are becoming harder to achieve, companies like Intel, AMD, TSMC, Samsung, and even Nvidia are devising tricks to increase efficiency. Gelsinger noted innovations like 3D packaging, gate-all-around transistors, backside power delivery, and lithography advances, echoing comments from AMD.

Hardware manufacturers have also begun adopting chiplet-based designs, which increase flexibility by allowing multiple semiconductor process nodes on a given product. Nvidia has become a primary proponent of AI and machine learning, introducing upscaling techniques to dramatically improve gaming performance.

Gelsinger declared that Intel hopes to progress from 100 billion transistors per package to 1 trillion before the end of this decade.

Permalink to story.

https://www.techspot.com/news/101330-intel-ceo-pat-gelsinger-moore-law-isnt-dead.html

 

[hahahanoobs]

I knew I heard that statement before....

AMD CEO Lisa Su says Moore's Law isn't dead, but has slowed down
May 2023

 

[Sherwoodnt]

So it’s still progressing at the same speed, but slower now. Got it.




?

 

[cristianm]

The law of diminishing returns is in full force for CPUs: double the transistors for at most 50% improvements. Combined with the slowing of Moore's law allows us to upgrade less often

 

[Theinsanegamer]

The law of diminishing returns is in full force for CPUs: double the transistors for at most 50% improvements. Combined with the slowing of Moore's law allows us to upgrade less often

AMD went from 4,800 million transistors in the 1800x to 4,150 million in the 5800x. A significant decrease. And it saw a 30-50% improvement over the 1800 depending on benchmark. 100% if you are talking memory latency (and anything memory heavy by extension).

Granted, the "10 years to obsolesence" rule has been around CPUs since the core 2 days. I enjoy getting my money's worth but I miss the days of rapid CPU improvements every year.

I knew I heard that statement before....

AMD CEO Lisa Su says Moore's Law isn't dead, but has slowed down
May 2023

Not satisfied with copying their "glue", intel has resorted to copying their news briefs as well!

 

[hahahanoobs]

AMD went from 4,800 million transistors in the 1800x to 4,150 million in the 5800x. A significant decrease. And it saw a 30-50% improvement over the 1800 depending on benchmark. 100% if you are talking memory latency (and anything memory heavy by extension).

Granted, the "10 years to obsolesence" rule has been around CPUs since the core 2 days. I enjoy getting my money's worth but I miss the days of rapid CPU improvements every year.

Not satisfied with copying their "glue", intel has resorted to copying their news briefs as well!

One could argue Intel glue (EMIB) and vertical compute tile chip stacking (Foveros) are superior. They do have a 47-tile Ponte Vecchio GPU already and vertical compute tile stacking on xonsumer parts like MTL.

 

[Watzupken]

“ building faster devices would theoretically require pumping more electricity into more transistors, significantly raising costs and energy consumption.”

Isn’t this already the case, especially for Intel? 3 years from Alder Lake to Raptor Lake refresh, just proves this point. Nvidia’s case may not be apparent, but because they jumped from some cheap Samsung matured node to a cutting edge TSMC node, there are significant improvements and benefits. Hence, Ada Lovelace looks very good. I doubt Blackwell will see significant improvements unless power requirement goes up.

 

[cristianm]

AMD went from 4,800 million transistors in the 1800x to 4,150 million in the 5800x. A significant decrease. And it saw a 30-50% improvement over the 1800 depending on benchmark. 100% if you are talking memory latency (and anything memory heavy by extension).

The 1800x was a monolithic CPU, whereas the 4.1B transistors of the 5800x are only for the compute die (to make it fair you should add the 2.1B transistors of the IO die of the 5800x). Since the IO die is not involved in computing, I believe a more accurate comparison is the compute die of the 3700x that has 3.8B transistors and the compute die of the 7700x that has 6,5B transistors.

 

[Mr Majestyk]

One could argue Intel glue (EMIB) and vertical compute tile chip stacking (Foveros) are superior. They do have a 47-tile Ponte Vecchio GPU already and vertical compute tile stacking on xonsumer parts like MTL.

Mentioning Ponte Vecchio doesn't help the argument. Rarer than hen's teeth and another massively delayed Intel failure.

 

[hahahanoobs]

Mentioning Ponte Vecchio doesn't help the argument. Rarer than hen's teeth and another massively delayed Intel failure.

You missed the point, which were the packaging differences....

 

[Gezzer]

Both Gelsinger and Haung are right. Because they're not referring to the same "law"

Haung is referring to the often quoted "transistor density per chip will double roughly every 2 years" This is a simplification of the original assertion/s by Moore, and yes for all intents and purposes it died quite a while ago IMHO. It's a media sound bite version of what Moore was discussing in his article and never really should of been considered a "law". Simply one of a number of observations of where current trends were heading at the time he wrote it.

Gelsinger OTOH is referring to the complete prediction/s that Moore's original article stated. Of which the often quoted one is simply a portion of the overall extrapolations. The entire article is essentially about how manufacturing processes and costs will continually improve to such a point that computing systems and the chips they use will become so powerful and inexpensive that computers will become ubiquitous in regards to our day to day uses and lives.

We've already seen the effects of Moore's predictions play out in real life. Smart phones are a perfect example. At one time the very idea that everyone could have access to a relatively inexpensive portable computing device that could rival previous super expensive "super computers" would of been laughed at by a lot of users. But it's true. And that prediction of Moore's is in fact still in effect, but is also slowing down as Gelsinger asserts due to diminishing returns.

In the end we will eventually hit a wall where diminishing returns will cause Moore's "law" to grind to a halt.. if we continue to use current technology. But it doesn't have to happen if we simply pivot to new technology such as quantum or optical computing. Both show varying degrees of promise depending on who you're listening to. But I'm fairly confident that Moore's over all predictions will hold true for a very long time. So Moore's law is dead, long live Moore's law... I guess?

Moore's original article:
https://hasler.ece.gatech.edu/Published_papers/Technology_overview/gordon_moore_1965_article.pdf

An in depth breakdown of what Moore was discussing:
https://arstechnica.com/gadgets/2008/09/moore/

A blog post on just how powerful current smart phones are:
https://blog.adobe.com/en/publish/2...ring-1980s-supercomputer-to-modern-smartphone