China is slowly moving up the microprocessing value chain
中国正在逐步向微处理价值链的上游攀登【技术季刊《中国的技术》系列之五】

由底至顶  From bottom to top-书迷号 shumihao.com

THE FORTUNE PRECISION EQUIPMENT COMPANY makes chunks of metal. Hulking sheets of it are cut with millimetre accuracy using robot arms in room-sized enclosures bearing the brand of their German or Japanese manufacturers. The white spray of cooling lubricant makes the process look like an industrial-strength shower for some post-modern Cleopatra.
富创精密设备公司生产大块金属。印有德国或日本制造商品牌的机柜有一个房间那么大,里面用机械臂在以毫米的精度切割大张金属板。冷却润滑剂的白色喷雾让这个过程看起来像是给某位后现代埃及艳后的工业级淋浴。

Based in Shenyang, five hours north-east of Beijing by train, Fortune is the bottom rung of the most important and complex supply chain on Earth: the one which produces the integrated circuits, or chips, found in smartphones and servers. Fortune’s robots make parts for equipment which will be installed in factories in Taiwan and Oregon, and used to etch circuitry on silicon and make chips. Selling equipment to industry giants like Applied Materials in America makes it a small success for the semiconductor supply chain. Much of the rest of the industry is not doing so well.
富创总部设在北京东北方向的沈阳,距离北京五小时火车车程。它是地球上最重要和最复杂的供应链中的最底层。这条供应链生产智能手机和服务器中的集成电路也就是芯片。富创的机器人生产的零部件会被嵌入台湾和俄勒冈州的工厂安装的设备中,这些设备在硅上蚀刻电路并制造芯片。把设备卖给美国的应用材料公司(Applied Materials)等行业巨头让它在半导体供应链上占有了一席之地。行业中许多其他公司的情况就没有这么好了。

Failure to make cutting-edge chips is not exactly China’s fault. It is a difficult industry to kick-start. The factories that produce the chips are phenomenally expensive. The technology itself is even more complex than an internal-combustion engine. The intellectual property behind cutting-edge processes is fiercely guarded. In many ways the manufacturing of chips represents the supreme technological challenge for China, an amalgamation of all the other challenges presented in this report. It will have to call upon everything it has learned from successes and failures like nuclear plants and engines if it is to succeed.
未能制造出最先进的芯片并不是中国的错。这是一个很难启动的行业。生产芯片的工厂耗资极大。这种技术本身甚至比内燃机更复杂。尖端流程背后的知识产权受到严密保护。在许多方面,芯片制造代表了中国面临的最高技术挑战,是本专题报道中提出的所有其他挑战的大融合。要想成功,中国就必须利用它从成功和失败(如核电站和发动机)中学到的一切。

The Chinese government is trying hard (the country’s biggest chip factory, SMIC, a private firm, has settled many suits over IP theft). In October the government raised 204bn yuan ($29bn) from the finance ministry, state-owned firms and local governments for its domestic chipmaking efforts. That followed 139bn yuan raised in 2014. The problem is that the government’s chip programme is optimising for the wrong thing. Instead of trying to stimulate a domestic chip industry to meet China’s huge market needs, the funds are being spent on trying to reach parity with chip companies like Intel.
中国政府非常努力。中国最大的芯片工厂、私营公司中芯国际(SMIC)已就许多知识产权盗窃案达成和解。2019年10月,中国政府从财政部、国有企业和地方政府筹集了2040亿元人民币,用于研发国产芯片。在此之前,2014年时已筹集了1390亿元人民币。问题在于政府芯片计划的优化目标选错了。这些资金并没有试图刺激国内芯片产业来满足中国巨大的市场需求,而是将资金用于追赶英特尔等芯片公司。

Chips are a vital product to China because they are fundamental to any technology-led growth that the country desires for its future, as well as for making weapons. PwC, a consultancy, estimates that the global market for chips will grow by 4.6% a year, to be worth $575bn in 2022, driven by the requirements of cars, AI systems and communications networks.
芯片对中国而言是至关重要的产品,因为它们对于中国渴望在未来实现的任何技术主导型增长以及武器制造来说都是一种基础。咨询公司普华永道估计,在汽车、人工智能系统和通信网络的需求推动下,全球芯片市场将以每年4.6%的速度增长,到2022年将达到5750亿美元。

Currently a huge share of that market value moves through China, but is not captured by it. The 418bn chips the country imported in 2018 cost $312bn, a quarter more than it spent importing crude oil. And beyond grabbing a larger part of the value chain for itself, controlling the production of chips would also give China indirect control over myriad other industries, from social networking to personal computing.
目前,这一市场价值中有极大一部分流经中国,但未被中国捕获。中国在2018年进口了4180亿枚芯片,花费3120亿美元,比在进口原油上的支出还多四分之一。除了把价值链中更大的一部分留给自己外,控制芯片的生产还可以使中国间接控制从社交网络到个人计算机的无数其他行业。

Most of the state-led efforts have failed so far. SMIC is on the verge of producing chips at levels of sophistication roughly equivalent to those reached by Intel a decade ago. Its revenues—$3.4bn in 2018—were about a tenth those of TSMC, its Taiwanese rival. SMIC is not yet globally known for its quality and reliability. But Fortune is making progress. It used to ship its metal components off to third parties in Japan and Taiwan to be cleaned up. Today it does not need to do that. The firm has its own clean rooms where it sands down its shiny aluminium components and gives them a smooth grey coating before vacuum packing them in thick plastic. The firm has also started shipping more complex components to its suppliers, simplifying what its customers have to do while capturing more of the value of the final product.
迄今为止,大多数国家主导的努力都以失败告终。中芯国际即将生产的芯片的成熟度大致与英特尔十年前的水平相当。它的收入(2018年为34亿美元)约为台湾竞争对手台积电(TSMC)的十分之一。中芯国际尚未以质量和可靠性享誉全球。但富创正在取得进展。它以前会将它生产的金属部件运送到日本和台湾的第三方清理。今天它不需要这样做了。公司拥有自己的洁净室,在其中打磨闪闪发亮的铝制部件,喷上光滑的灰色涂层,然后再用厚塑料真空包装。它还开始向供应商供应更复杂的元件,从而简化了客户要做的处理,同时捕获了最终产品中更多的价值。

Although Chinese firms are still behind in the manufacture of chips, they have recently achieved some success with designing them for AI applications. In late September the nerdier corners of Silicon Valley were abuzz after Alibaba, a Chinese tech-giant, released Hanguang 800, a chip designed specifically for carrying out machine-learning tasks. Even though Alibaba relied on TSMC in Taiwan to fabricate the chips, AI engineers in the Valley remarked on the Hanguang’s performance, stating that it had beaten all other chips in its class. This was not supposed to happen, as China was thought to be well behind American chip companies.
尽管中国公司在芯片制造方面仍然落后,但它们最近为人工智能应用设计芯片方面取得了一些成功。去年9月下旬,中国技术巨头阿里巴巴发布了专门用于执行机器学习任务的芯片“含光800”,让硅谷的书呆子们议论纷纷。尽管阿里巴巴依靠台湾台积电制造芯片,但硅谷的人工智能工程师还是对含光的性能发表了评论,称其击败了同级别的所有其他芯片。这真是出人意料,因为人们认为中国远远落后于美国芯片公司。

Faster, higher, stronger
更快更高更强

On November 6th the latest results of MLPerf, an industry-standard benchmark for AI chips, were published. They showed that the Hanguang 800 chip was performing a standardised machine-learning task 13 times faster than the chip that Intel had just released (see chart). The comparison is not totally fair, as the Alibaba chip was made physically larger than the Intel chip, letting it draw more power and perform more calculations per second. But even compared with a bigger chip from Nvidia called the Titan RTX, the Hanguang 800 clocked in four times faster.
去年11月6日,人工智能芯片的行业标准化基准MLPerf发布了最新测试结果。它表明,含光800芯片执行标准化的机器学习任务的速度是英特尔刚刚发布的芯片的13倍(见图)。这种比较并不完全公平,因为阿里巴巴芯片的物理尺寸要比英特尔的芯片大,这让它的功耗更高,每秒执行更多的计算。但即使与英伟达尺寸更大的芯片Titan RTX相比,含光800的处理速度也要快上四倍。

There are probably more caveats. Alibaba ran only one out of five tests. Poor performance in the others would betray a chip over-optimised for one task. But even in the most pessimistic scenario it is impressive. That a Chinese company has designed an AI chip which performs as well as, or better than, its Western competitors should alert American politicians and innovators to China’s progress in this area.
它很可能还有更多不足。阿里巴巴只进行了五分之一的测试。在其他测试中表现不佳会揭示出一种芯片针对某项任务过度优化了。但哪怕出现最悲观的情况,也足以令人印象深刻了。中国公司设计的人工智能芯片的性能相当于甚至优于西方竞争对手,这应该使美国的政客和创新者警惕中国在这一领域的进步。

Whereas Fortune’s more industrial flavour of success in the manufacturing supply chain took a traditional route for Chinese firms—start at the bottom and work up—Alibaba’s success in designing a chip is more interesting. It is rooted in its wide deployment of machine-learning systems across its business, both in its Taobao online market and its Hema shops on the high street. The firm processes billions of images a day as part of its normal operations, and the machine-learning software it has trained to do that work is now very accurate and powerful.
富创在制造业供应链中偏工业化的成功模式走的是中国公司的传统路线——从最底层开始逐渐向上游走,而阿里巴巴在芯片设计方面的成功更耐人寻味。它植根于广泛部署在公司所有业务(包括网上市场淘宝和商业街上的盒马鲜生店)中的机器学习系统。作为日常运营的一部分,这家公司每天要处理数十亿张图片,而它经过训练来做这件事的机器学习软件现在已经非常准确且强大。

The Hanguang 800’s designers spent a lot of time sitting next to the coders who built those algorithms. Their job was to work out how to render the algorithms in silicon, so the more time they could spend learning from engineers writing high-performance algorithms, the better. By being close to the market in which AI is used, like Taobao’s and Alibaba’s offline shops, the Hanguang designers were able to tweak the design of the chip to optimise its performance on those tasks. Indeed, in many ways, the data-labelling grunt work that makes Alibaba’s machine-learning algorithms as good as they are translates directly into the high performance of its new chip. China’s strength in data-labelling at the very bottom of the AI supply chain is translating into design strength at the top.
含光800的设计师花了很多时间坐在构建这些算法的程序员身边。他们的工作是弄清楚如何把算法搬上硅片,因此花在向工程师学习编写高性能算法上的时间越多越好。由于靠近淘宝和阿里巴巴的线下商店等使用人工智能的市场,含光的设计人员能够调整芯片的设计来优化它在这些任务上的性能。确实,从许多方面来说,数据标注的艰巨工作让阿里巴巴的机器学习算法达到了今天的出色水准,也直接转化成了新芯片的高性能。中国在人工智能供应链的最底层——数据标注——方面的优势正在转化为顶层的设计优势。

Must do better
必须做得更好

In other parts of the semiconductor supply chain, things are less rosy. Several executives, who asked for their names not to be attached to criticism of government policy, said that the stimulus had been going to the wrong place in trying to catch up with Western giants such as Intel, or TSMC in Taiwan. While Hanguang 800 is remarkable, Alibaba’s design work is a far less capital-intensive, less complex process than the physical manufacture of a chip (which, in Hanguang’s case, was still done in Taiwan).
在半导体供应链的其他部分,情况则没那么乐观。几位高管(要求在批评政府政策时隐去姓名)表示,试图赶上英特尔这样的西方巨头或台湾的台积电的刺激措施是用错了地方。尽管“含光800”非常出色,但相比芯片的物理制造(含光依然是在台湾制造的),阿里巴巴的设计工作的资本密集度和复杂度要低得多。

If catching up on internal-combustion engines was hard, doing so on traditional semiconductor manufacturing will be close to impossible. The market for chips is changing fast. Instead the government would do well to focus on stimulating both design and manufacture of chips aimed at middle-end markets such as the internet of things, and emerging areas like AI.
如果在内燃机上实现赶超很难,那么在传统半导体制造上就几乎是不可能的。芯片市场瞬息万变。相反,政府更好的做法是专注于刺激针对物联网等中端市场以及人工智能等新兴领域的芯片设计和制造。

In one way, China’s challenge with chips is even harder than the problem it faced with combustion-engine cars. There is no completely new technology arriving which will free China from the need to catch up with the rest of the world. But China’s chip companies should still listen to the market, not chase prestige.
在一个方面,中国在芯片方面的困难甚至比内燃机汽车面对的问题还要艰巨。在芯片领域还没有哪种全新的技术出现,可以让中国无需赶超世界其他地区。但中国的芯片公司仍然应该倾听市场,而不是追逐虚名。

Manufacturing and designing chips for the internet of things and AI applications offers an opportunity to leapfrog less agile chip companies, even if it is not the wholesale opportunity that electric vehicles present. If they can focus on these new areas and be patient, it is likely that the scale and depth of the country’s resources could end up winning its companies a permanent spot high on the global supply chain for semiconductors.
制造和设计用于物联网和人工智能应用的芯片为中国提供了一个机会来超越那些不太敏捷的芯片企业,哪怕这个机会不像在电动汽车领域那样俯拾即是。如果它能够专注于这些新领域并保持耐心,那么以中国资源的规模和深度,很可能最终为它的企业赢得全球半导体供应链上的永久高点。