來源：內(nèi)容來自財富，作者臺積電劉德音

    半個多世紀以來，半導體一直是技術(shù)創(chuàng)新的核心，技術(shù)的進步與半導體性能、能耗和成本的發(fā)展同步?，F(xiàn)在，隨著對高性能計算 (HPC) 以及 5G 和 AI 應(yīng)用的需求不斷增長，對技術(shù)進步的需求猛增，為半導體技術(shù)新想象的未來鋪平了道路，其中可以擁有無限可能實現(xiàn)。

    為了理解這個未來，回顧過去 60 年是有意義的，當時人們發(fā)明了一種將許多晶體管放在同一個芯片上的方法——集成電路 (IC) 或微芯片。在隨后的幾年中，半導體技術(shù)通過不斷的小型化而進步，這包括按照摩爾定律預(yù)測的那樣，集成電路上的晶體管數(shù)量每隔一年就會翻一番，該定律以美國工程師戈登·摩爾的名字命名。這種持續(xù)的進步使我們的手機擁有比 1969 年阿波羅 11 號登上月球的現(xiàn)在古老的 70 磅重的計算機更強大的計算能力。

    半導體技術(shù)和集成電路的一個關(guān)鍵屬性是不斷降低每個功能的成本。隨著時間的推移，這種持續(xù)的成本降低導致了半導體技術(shù)的普遍部署。例如，可視電話于 1970 年由 AT&T 首次商業(yè)化，但由于成本高昂，它的客戶不到 500 個。

    35年前臺積電首創(chuàng)的純代工模式的誕生，幫助半導體技術(shù)大規(guī)模降低成本。在這種模式下，純代工廠運營的半導體制造廠專注于為其他公司生產(chǎn) IC，而不是提供自己設(shè)計的 IC 產(chǎn)品。由于 IC 生產(chǎn)設(shè)施的建造和維護成本高昂，并且可能會大量消耗公司的資金，因此將這種生產(chǎn)外包給代工廠可以讓公司將資源集中在最終產(chǎn)品上。這使得無晶圓廠（僅限設(shè)計）行業(yè)蓬勃發(fā)展，并幫助實現(xiàn)了使遠程工作、在線學習、共享經(jīng)濟和娛樂流媒體成為現(xiàn)實的技術(shù)的大規(guī)模無處不在的部署。

    COVID-19 及其帶來的封鎖成為技術(shù)創(chuàng)新的另一個轉(zhuǎn)折點，在一年內(nèi)發(fā)生了超過 10 年的數(shù)字化，增加了對半導體的需求。根據(jù)麥肯錫公司的數(shù)據(jù)，按照目前的速度，到 2030 年，全球半導體年收入將增長到超過 1 萬億美元，直接貢獻 3 萬億至 4 萬億美元的全球電子產(chǎn)品增長。然而，持續(xù)降低成本的承諾導致人們低估了半導體的價值。正如最近的半導體供應(yīng)鏈挑戰(zhàn)清楚地表明，半導體無處不在，在現(xiàn)代社會中發(fā)揮著寶貴而重要的作用。

    隨著計算設(shè)備變得無處不在，通過全球網(wǎng)絡(luò)生成和通信的數(shù)據(jù)量（通常是實時的）呈指數(shù)級增長。為了跟上這種增長，高性能計算 (HPC) 變得至關(guān)重要，并且正在呈現(xiàn)爆炸式增長。HPC 是高速處理數(shù)據(jù)和執(zhí)行復雜計算以解決性能密集型問題的能力。如今，HPC 已經(jīng)超越智能手機成為增長動力。根據(jù)Report Ocean的研究，它是半導體行業(yè)增長最快的領(lǐng)域之一，預(yù)計到 2027 年，全球 HPC 芯片組市場規(guī)模將從 2019 年的 43 億美元達到 136.8 億美元。

    虛擬世界與物理世界的融合將給社會互動的方式帶來翻天覆地的變化，并將通過 HPC 應(yīng)用程序?qū)崿F(xiàn)。除了由半導體制成的大量傳感器和執(zhí)行器之外，虛擬世界和物理世界的這種集成還需要智能設(shè)備、可穿戴設(shè)備、物聯(lián)網(wǎng)等硬件，以及 5G、人工智能和大數(shù)據(jù)分析等技術(shù)，用于交流和理解信息, 和決策。對于這些應(yīng)用中的每一個，半導體含量及其提供的價值都將迅速增加。

    半導體將為越來越多的產(chǎn)品注入智能和新功能，從而提升這些產(chǎn)品的價值。例如，自動駕駛汽車將通過先進的芯片變得更加安全和節(jié)能，這些芯片允許執(zhí)行復雜的軟件功能和分析。德克薩斯大學的研究估計凈能源減少 11% 至 55% 與美國當前的地面交通條件相比，基于這種預(yù)期的自動駕駛汽車能源效率。社會也期待著超出我們今天想象的新用戶應(yīng)用程序。半導體提供的計算能力將推動個性化和社區(qū)醫(yī)學以及疫苗和藥物發(fā)現(xiàn)。打擊社交媒體上的虛假信息需要更好的算法和計算能力來訓練人工智能模型。

    例如，用于創(chuàng)建逼真的人類質(zhì)量文本的最先進的 AI 語言模型之一 GPT-3需要 300 zetta-FLOPS（超級計算機性能的衡量標準）才能在高性能計算云上進行訓練。作為回報，這種 AI 語言模型所支持的能力令人印象深刻。GPT-3 最近被《紐約時報》的科技專欄作家 Kevin Roose 用來完成書評。

     人工智能通常被認為是一種主要涉及軟件和算法的技術(shù)。然而，硬件技術(shù)打開了通往虛擬世界的大門，讓我們能夠使用從人工智能中獲得的信息。因此，即使在虛擬世界中，物理也占據(jù)了中心位置。

    隨著半導體技術(shù)的進步以滿足 5G 和 AI 時代的需求，能源效率已成為最重要的指標，不僅因為計算能力已經(jīng)因無法散熱而受到限制，而且因為全球計算能源使用的升級速度比任何其他應(yīng)用領(lǐng)域。僅由于半導體技術(shù)，計算的能源效率一直在快速發(fā)展——每兩年提高 2 倍——人們普遍樂觀地認為，技術(shù)將像過去 50 年那樣繼續(xù)像發(fā)條一樣發(fā)展。

    這種經(jīng)常與摩爾定律混為一談的樂觀主義也許比“定律”本身更重要。正是這種行業(yè)和整個社會共同的樂觀態(tài)度，推動了行業(yè)迎接挑戰(zhàn)，并使預(yù)言成為自我實現(xiàn)的預(yù)言。

    在接下來的 50 年中，下一代可能會使用虛擬現(xiàn)實和增強現(xiàn)實 (VR/AR) 作為他們與世界互動的主要方式。今天的 VR/AR 頭顯平均重量超過一磅，電池壽命不到兩到三個小時，而且價格高昂，這讓我們想起了 25 年前的手機。要達到與當今手機相同的普及水平，VR/AR 設(shè)備需要提高 100 倍以上。這只能通過不斷進步的半導體技術(shù)來實現(xiàn)。

未來幾十年將是半導體行業(yè)的黃金時代。

    在過去的 50 年里，半導體技術(shù)的發(fā)展就像在隧道里行走。前進的道路很明確，因為每個人都在努力遵循一條明確的道路——縮小晶體管?，F(xiàn)在我們正在接近隧道的出口。隧道之外還有更多的可能性：從材料到架構(gòu)的創(chuàng)新使新路徑成為可能，以及由新應(yīng)用定義的新目的地。我們不再受隧道的限制，我們現(xiàn)在擁有無限的創(chuàng)新空間。

附原文：

TSMC chairman Mark Liu describes how the world’s largest chipmaker is reimagining the semiconductor industry

——by Mark Liu

     For over half a century, semiconductors have been at the heart of technological innovation, with advancements in technology marching to the cadence of developments in semiconductor performance, energy consumption, and cost. Now, with the ever-growing demand for high-performance computing (HPC), as well as 5G and A.I. applications, the need for technological advancement has skyrocketed, paving the way for a newly imagined future for semiconductor technology, where infinite possibilities can be realized.

   To understand this future, it makes sense to look back 60 years in the past, to the invention of a way to put many transistors together on the same chip—the integrated circuit (IC) or microchip. Throughout the years that followed, semiconductor technology advanced through continuous miniaturization, which involved doubling the number of transistors on an integrated circuit every other year as predicted by Moore’s law, named after American engineer Gordon Moore. This continued advancement is what allows our mobile phones to have far more compute power than the now ancient 70-pound computer that landed Apollo 11 on the moon in 1969.

    A key attribute of semiconductor technology and the integrated circuit has been relentless reduction of cost per function. This continuous cost reduction led to ubiquitous deployment of semiconductor technologies over time. The picture-phone, for instance, was first commercialized in 1970 by AT&T, but because of its high cost, it had fewer than 500 customers.

    Large-scale cost reduction of semiconductor technology was helped along by the birth of the pure-play foundry model, pioneered by TSMC at its establishment 35 years ago. In this model, pure-play foundries operate semiconductor fabrication plants focused on producing ICs for other companies instead of offering IC products of their own design. As IC production facilities are expensive to build and maintain, and can be a huge drain on finances for companies, outsourcing this production to foundries allowed companies to focus their resources on their end product. This allowed the fabless (design only) industry to flourish and helped enable the large-scale ubiquitous deployment of the technologies that make remote working, online learning, the sharing economy, and entertainment streaming a reality today.

    COVID-19 and the lockdowns it brought along with it became another turning point for technology innovation with more than 10 years’ worth of digitization happening over a single year, increasing the demand for semiconductors. At the current pace, annual global semiconductor revenue will grow to more than $1 trillion by 2030, directly contributing to $3 trillion to $4 trillion of global electronics growth, according to McKinsey & Co. Yet, the promise of continuous cost reduction has created an expectation that underestimates the value of semiconductors. As the recent semiconductor supply-chain challenge so clearly illustrates, semiconductors are everywhere and fulfill a valuable and vital role in modern society.

    As computing devices become ubiquitous, the amount of data generated and communicated across a global network, often in real time, has grown exponentially. To keep up with this growth, high-performance computing (HPC) has become crucial and is seeing explosive growth. HPC is the ability to process data and perform complex calculations at high speeds to solve performance-intensive problems. Today, HPC has already surpassed the smartphone as a growth driver. It is one of the fastest growing segments of the semiconductor industry, with the global HPC chipset market size expected to reach $13.68 billion by 2027 from $4.30 billion in 2019, according to research from Report Ocean.

    The integration of the virtual with the physical world will bring about a sea change in the way society interacts with one another and will be enabled by HPC applications. In addition to the multitude of sensors and actuators made of semiconductors, this integration of the virtual and the physical worlds requires hardware like smart appliances, wearable devices, IoT, and technologies like 5G, A.I., and big-data analytics for communicating, understanding information, and decision-making. For each of these applications, the semiconductor content, and the value it provides, will increase rapidly.

   Semiconductors will imbue intelligence and new functionalities into more and more products, elevating the value of such products. For example, autonomous driving vehicles will become even safer and more energy efficient with advanced chips which allow for the execution of complex software functionalities and analytics. University of Texas research estimates a net energy reduction of 11% to 55% versus the current ground transportation conditions in the U.S., based off this expected autonomous vehicle energy efficiency. Society is also expecting new user applications beyond what we can imagine today. Personalized and community medicine as well as vaccine and drug discovery will get a boost from the computing power provided by semiconductors. Combating disinformation on social media will need better algorithms and computing power for training A.I. models.

    As an example, one of the most advanced A.I. language models for creating realistic human-quality text, the GPT-3, requires 300 zetta-FLOPS (a measure of supercomputer performance) to train on a high-performance compute cloud. In return, the capability enabled by this A.I. language model can be impressive. GPT-3 recently was used by Kevin Roose, a tech columnist for the New York Times, to complete a book review.

A.I. is often thought of as a technology involving primarily software and algorithms. Yet, hardware technology is what opens the door to the virtual world and allows us to use the information derived from A.I. Thus, even in the metaverse, the physical takes center stage.

    As semiconductor technology advances to meet the needs of the 5G and A.I. era, energy efficiency has become the most important metric not only because computing power is already throttled by the inability to remove heat, but also because the global energy use of computing escalates faster than any other application area. Energy efficiency of computing due to semiconductor technology alone has been advancing at a rapid pace—2X every two years—and there is shared optimism that technology will continue to advance like clockwork as it did over the past 50 years.

    This shared optimism that is often conflated with Moore’s law is perhaps more important than the “law” itself. It is this shared optimism by the industry and society at large, that has propelled the industry to meet the challenge and make the prophecy a self-fulfilling one.

    In the next 50 years, the future generation will likely use virtual- and augmented-reality (VR/AR) as their principal means of interaction with the world. Today’s average VR/AR headsets weigh well over a pound, with a battery life of less than two to three hours, and a high price tag, which reminds us of the cell phones of 25 years ago. To achieve the same level of ubiquity as today’s cell phones, VR/AR devices will need to improve by more than 100 times. This can only be done with continuous advancement of semiconductor technology.

    The upcoming decades will be a golden era for the semiconductor industry. Over the past 50 years, the development of semiconductor technology has been akin to walking inside a tunnel. The way ahead was clear as there was a well-defined path that everyone diligently followed—shrinking the transistor. Now we are approaching the exit of the tunnel. There are many more possibilities outside the tunnel: new paths made possible by innovations from materials to architecture and new destinations defined by new applications. We are no longer bound by the confines of the tunnel, and we now have unlimited room for unleashed innovation.

免責聲明： 本文章轉(zhuǎn)自其它平臺，并不代表本站觀點及立場。若有侵權(quán)或異議，請聯(lián)系我們刪除。謝謝！     矽源特科技ChipSourceTek