news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

in april 1965, electronics magazine published an article by intel co-founder gordon moore, which pointed out that the number of transistors that can be accommodated on an integrated circuit will double approximately every two years.

the article and its predictions have since become legend, and like most legends, it has undergone many changes in telling and retelling. the media seized on the article’s argument that semiconductor technology would usher in a new era of electronic integration and distilled it into a maxim that has been presented in many forms over the years. however, no matter how it is presented, it is always given the same name: moore’s law.

who is the most loyal guardian of moore's law? if this question was asked ten years ago, the answer was almost unique. apart from intel, co-founded by gordon moore, who else could take on this responsibility?

since its founding in 1968, intel has been closely associated with the word “shrink.” for its first four decades, it was high praise.about every two years, the u.s. chip pioneer would introduce a new transistor that was half the size of the previous one, allowing more chips to fit on nearly the same silicon wafer and sell profitably at nearly the same price. that enabled intel to dominate the market for memory chips and, once “memory” became commoditized in the 1980s, the market for the microprocessors that powered the pc revolution.

but ten years later, the answer is no longer unique. there are fewer and fewer believers in moore's law, and intel is no longer a pioneer in the chip industry. after two consecutive disastrous quarterly financial reports, intel's market value has shrunk from more than $210 billion in january to $84 billion, even lower than the value of its factories and equipment. it can be said that in the company's decades of history, there has never been such a difficult time.

a new question arises: who can truly inherit moore's law?

nvidia

nvidia seems to have always disagreed with moore's law proposed by the founder of intel.

as early as 2010, bill dally, chief scientist and senior vice president of research at nvidia, emphasized in a column in forbes that moore's law was dead. he said that parallel computing was the future of the semiconductor industry.

he believes that the cpu shrinking predicted by moore's law is now no longer true, and cpu performance no longer doubles every 18 months. this poses a serious threat to many industries that rely on the historical growth of computing performance:

public agencies need more computing power to predict dangerous weather events and analyze long-term climate change. energy companies need to evaluate large amounts of seismic and geological data to find new ways to safely extract oil and gas from existing reserves. pharmaceutical researchers need increased computing power to design drug molecules that bind to specific cell receptors. clinical oncologists need better and faster medical imaging to diagnose cancer and determine treatments. heart surgeons want to visually assess damaged tissue in real time to ensure their operations are working.

he said the industry urgently needed the emergence of parallel computers. one of its basic advantages is that it can effectively convert more transistors into higher performance. doubling the number of processors can double the speed of many programs. in contrast, doubling the number of transistors in a serial cpu can only bring very limited performance improvements-and the energy consumption is very high.

more importantly, parallel computers (such as graphics processing units, or gpus) are able to continue to scale computing performance in today's energy-constrained environment. every three years, nvidia can quadruple the number of transistors (and cores). by running each core slightly slower, thereby increasing efficiency, nvidia can more than triple performance for the same total power.

to put it simply, the cpu represented by intel is no longer the future of computing, and the gpu represented by nvidia is the real future. at the time, it was a very radical and interesting point of view.

many years later, at the 2018 gpu technology conference (gtc) held in san jose, california, nvidia ceo huang renxun took the stage and once again denied moore's law in his own way. he repeatedly emphasized that due to the extreme advancement of technology, gpus are following a law of their own.

"now there is a new law," he said, "the law of super acceleration." this law was later recognized as "huang's law."

how fast is gpu technology advancing?in huang’s keynote, he noted that nvidia’s gpus are 25 times faster today than they were five years ago. if moore’s law had held, they would only have gotten 10 times faster.

he then illustrated the gpu performance gains with another benchmark: the time it takes to train alexnet, a neural network trained on 15 million images. five years ago, he said, that process took six days to complete using two nvidia gtx 580 graphics cards; now it takes just 18 minutes using the company’s latest hardware, the dgx-2 — 500 times faster.

huang made it clear that gpus need their own law because they benefit from simultaneous advances in many areas: architecture, interconnect, storage technology, algorithms, etc. "innovation is not just in the chip," he said, "but in the entire technology stack.

since 2018, huang has reiterated his view that moore's law has reached its end, frequently referring to the concept as "dead" and telling wired magazine earlier this year: "we have to throw away moore's law so that we can think about new ways to scale."

here comes the interesting thing. huang's law, which he strongly advocates, is not as beautiful as described. it is even a relatively abstract concept. huang renxun did not propose a law with specific numbers to follow. bill dally, chief scientist and senior vice president of research at nvidia, later stated that from november 2012 to may 2020, the performance of nvidia chips in the important field of ai computing has increased by 317 times, an average of more than doubling per year.open ai says that based on classic ai image recognition tests, performance doubles approximately every year and a half.

in short, huang's law is faster than moore's law, which doubles every two years, in nvidia's propaganda. even when discussing the future development of data centers recently, huang renxun reiterated the view that moore's law has ended. he believes that in the next 10 years, modern data centers will accelerate development and become more dense and energy-efficient. the traditional doubling speed every five years has ended, and the doubling every 10 years that everyone expects is becoming more and more difficult.

moore's law has indeed slowed down, but whether huang's law can replace it remains a big question mark.

in fact, most of nvidia's performance improvements are closely related to node switching.nvidia has twice recently delivered significant performance increases without a node switch - the first was from kepler to maxwell, and the second was from volta to turing.while nvidia is good at squeezing additional performance from the same node, we can also see the importance of new process nodes to nvidia's overall performance.

in fact, the so-called huang's law is driven to a considerable extent by moore's law. without the improvement of transistor density, there will be no more powerful gpu performance. if moore's law runs into trouble - whether in transistor scaling or in poorly defined performance improvements, huang's law will also run into trouble. as the benefits of node conversion decline, the speed of ai performance improvement will also slow down, which is why huang renxun and nvidia themselves have rarely mentioned their own law in the past two years.

in addition, moore's law does not just represent the continuous improvement of transistor density, its greater significance lies in obtaining chips with more powerful performance at a lower price.simply put, if the same specifications of ics are produced on wafers of the same area, with the advancement of process technology, ic output can double every year and a half. converted into cost, this means that the cost can be reduced by 50% every year and a half, and the average annual cost can be reduced by more than 30%. this is one of the keys to the continued prosperity of the consumer electronics market.

what about nvidia's huang's law? nonprofit research organization epoch found that between 2006 and 2021, gpu price/performance (measured in flops/$) doubled every 2.5 years, much slower than huang's law had previously predicted.

using a dataset of 470 graphics processing unit (gpu) models released between 2006 and 2021, epoch found that floating point operations per second per dollar (hereafter referred to as flop/s per dollar) doubled approximately every 2.5 years. for the top gpus at any point in time, the rate of improvement was found to be slower (doubling every 2.95 years for flop/s per dollar), while for gpu models commonly used in ml research, the rate of improvement was found to be faster (doubling every 2.07 years for flop/s per dollar).

from this perspective,huang's law may be good news for the ai ​​industry, but it is very different from moore's law, which benefits the entire semiconductor industry. nvidia is not the company that inherited moore's law.

tsmc

coincidentally, tsmc, which manufactures chips for nvidia, has shown great interest in continuing moore's law.

as early as 1998, tsmc chairman morris chang said that moore's law has been quite effective in the past 30 years and should still be applicable in the next 10 to 15 years. he said that the characteristics of the semiconductor industry are "mountains and rivers are so steep that it seems there is no way out", but the truth is that "there is always light after the dark clouds."

at the 2019 hot chips conference, tsmc research vice president philip wong emphasized in his speech that moore's law is not only still valid, but will remain valid for the next thirty years as long as the right technical means are adopted.

“it’s not dying,” he told attendees, “it’s not slowing down, it’s not even sick.”

in wong's view, the only important factor in maintaining moore's law is to keep increasing density. although he acknowledged that clock speeds have plateaued with the demise of dennard scaling, transistor density will lead to better performance and energy efficiency.

according to wong, as long as the company can continue to pack more transistors into a smaller space and improve energy efficiency, that's all that matters. in the short term, this will most likely be achieved in the traditional way, by improving cmos process technology to make transistors with smaller gate lengths.

in the long run, the slowdown in 2d scaling does not mean the end of density. he pointed out that even after the end of dennard scaling, there have been many innovations in semiconductor manufacturing that have kept density on an upward trend, in particular, the use of strained silicon and high-k metal gate technology, followed by the introduction of 3d finfet structures.

wong also emphasized the importance of packaging for the continuation of moore's law. in the short term, 2.5d packaging can be relied on (in fact, it is already in large-scale use today), while in the long term, true 3d packaging technology will lead to continued improvement in chip density.

at philip wong, tsmc has emphasized its own proposition more than once: moore's law is a symbol. as a foundry, it may not fully follow this law, but tsmc has been pursuing the direction indicated by this law.

in a july 2024 interview with ian cutress of techtechpotato, dr. kevin zhang, tsmc's head of process technology, said he didn't care about moore's law as long as overall progress was going well.

“well, my answer is simple — i don’t care,” zhang said. “as long as we can continue to push technology scaling, i don’t care if moore’s law still exists.”

“[observers] have defined moore’s law narrowly based on two-dimensional scaling, but that’s no longer the case,” he said. “if you look at the hype around innovation in our industry, we actually continue to find different ways to pack more functionality and capabilities into smaller form factors. we continue to achieve higher performance and higher energy efficiency. so from that perspective, i think moore’s law or technology scaling will continue.”

when asked about tsmc's success in incremental process node improvements, he clarified that their progress is by no means insignificant. tsmc emphasized that the foundry's transition from 5nm to 3nm-class process nodes has resulted in more than 30% improvement in ppa per generation, and will continue to make smaller but continuous improvements between major nodes in the future to enable customers to benefit from each new generation of new technology.

tsmc and nvidia have two completely different attitudes towards moore's law. the former tries its best to continue it, while the latter keeps denying it.so who is right? let's first take a look at what the inventor of moore's law originally said.

intel

unlike tsmc, which is on the rise, intel has indeed been on a downward trend in the past two years, but it seems to have never given up trying to continue moore's law.

since taking office as ceo in 2021, intel ceo gelsinger has always emphasized that moore's law is "still healthy." he even said that intel can exceed the speed of moore's law at least until 2031 and promote "super moore's law", which is to use 2.5d and 3d chip packaging technologies such as foveros to increase the number of transistors.

in december 2023, during a speech at mit, kissinger was asked about the possible end of moore's law. he said, "we are no longer in the golden age of moore's law. it is much more difficult now, so we may double every three years now, so we are definitely seeing a slowdown."

kissinger also stressed that although moore's law seems to be slowing down, intel will still be able to produce chips with 1 trillion transistors by 2030. today, the largest chip in a single package has about 100 billion transistors.the ceo said four factors made that possible: new ribbonfet transistors, powervia power delivery, next-generation process nodes and 3d chip stacking. he closed his answer by saying, “to all the critics who are declaring us dead… we are not done until the periodic table runs out.”

and at this year’s computex, intel’s keynote opened with ai co-founder gordon moore repeating one of his most famous quotes: “anything that has been done can probably be surpassed.” this has been intel’s battle cry in its pursuit of microchip dominance and one of the symbols of moore’s law.

in his speech, kissinger recalled the introduction of the intel 80286 in 1981, a 16-bit microprocessor containing 100,000 transistors, and stressed that modern chips are seeking to accommodate a billion transistors, a number expected to reach a trillion by 2020.

"contrary to what jen-hsun huang would have you believe, moore's law still holds true," gelsinger said, stressing that as the leading pc chip provider, intel will play an important role in the popularization of ai.

if tsmc is just doing its best to follow moore's law, intel seems to be risking its life and property for it.

although intel has performed very poorly in the past two quarters, and even outsourced some of its chips to tsmc, it has not given up its original vision and once again emphasized the 18a node that has appeared countless times in its technology roadmap.

let's review the technical details. 18a is the fifth production process in intel's roadmap to accelerate its return to technological leadership. intel 7 has appeared in alder lake and raptor lake cpus, while intel 4 just came out at the end of last year with meteor lake chips. the abandoned 20a was originally planned to be launched with the arrow lake cpu series, and the 18a node is expected to be officially launched in 2025.

it is understood that the first products using 18a process technology will be client pc processors code-named panther lake and data center processors code-named clearwater forest. in addition, intel's cpu code-named diamond rapids will also use this node. among external customers, microsoft confirmed plans to use it for processors, the us department of defense will also use it in its chips, and intel expects to have 8 18a tapeouts by mid-2025, including internal and external products.

"i'm betting the entire company on 18a." kissinger said everything depends on intel's 18a process, which is the company's biggest bet ever because it also puts tremendous pressure on intel's finances.

in fact, we have already seen the price intel paid for the 18a node: poor quarterly financial reports, plummeting stock prices. not to mention nvidia, even colgate, a toothpaste company in the united states, has a higher market value than intel.

in the eyes of many people, intel is now a bruised and battered boxer, and 18a is its last punch.

but fortunately, the original guardian of moore's law still wants to continue it rather than deny it.

final thoughts

who is the true successor to moore's law? it is unlikely that the answer to this question will be available in the short term. we also cannot immediately answer which of nvidia, tsmc and intel is the right party.

but what is certain is that only those with the most steadfast belief will have the last laugh, just like when intel decisively gave up memory and plunged into the microprocessor track, the only thing the semiconductor industry can believe in is truly advanced technology.