https://mil.news.sina.com.cn/dgby/2019-06-25/doc-ihytcerk9077535.shtml
美国的最新禁运 真的会让中国下代超算遭致命打击？
美国的最新禁运 真的会让中国下代超算遭致命打击？
588
最近这几天，有许多美国媒体幸灾乐祸，认为中国超算行业将受到致命打击，离开了美国的CPU，中国将造不出超算。事实是这样的吗？这个禁令对我国的超算行业有什么影响？
据多家媒体报道，在当地时间6月21日，美国商务部以国家安全为由，再次将中国的5家实体列入进出口管制的实体清单，禁止它们从美国的供应商采购零部件，相关决定于6月24日生效。这五家实体与中国超算行业的“三巨头”中的两位相关，涉及到“曙光”和“神威”两款超算。这五家实体中的江南计算技术研究所就是“神威”超算的研发制造单位。而其他四家实体，中科曙光，天津海光（Higon）、成都海光集成电路、成都海光微电子技术全都与“曙光”系列超算有关。
最近这几天，有许多美国媒体幸灾乐祸，认为中国超算行业将受到致命打击，离开了美国的CPU，中国将造不出超算。事实是这样的吗？这个禁令对我国的超算行业有什么影响？本文将对这个问题进行详细分析。
首先，美国为什么要再次对我国超算企业进行制裁？
实际上，早在2015年初，美国商务部就将中国的四大超算中心列入了禁运清单。这源于美国长久以来的焦虑：在大约40年中，中国的超算行业从完全进口美国产品，到购买美国的CPU自己制造超算，再到使用自己的CPU和加速器制造超算； 从只有几台超算，在全世界超算市场里不值一提，到拥有超算的数量全球第一，大大超越美国；从拥有的超算性能低下排不上号，到称霸全球最快超算排行榜第一名多年；中国超算行业走出了一条完美的逆袭之路。
世界超算组织top500.org每半年推出一次全球最快的500台超算排行榜。今年6月刚刚推出了最新的榜单。全球最快的500台超算数量按国家的分布图如下：
▲数据来源于top500.rog，由笔者整理
TOP500榜单中，中国超算的数量达到了219台，占全球的44%，世界第一；美国以116台的数量占全球的23%，比中国落后甚多。
要知道，在仅仅一年前的2018年6月，中国在TOP500中的超算总数还是206台，美国还是124台。这一年中，中国又有了进步，美国又在退步。从趋势来看，中国的优势会越来越大。
由于超算在军事，科研和经济中的巨大作用，美国心里的焦虑是可以理解的。他们急于抑制住中国的发展势头，于是就把商业诚信和公平贸易原则仍进下水道，祭出了禁运大法。那么，这次的禁运措施对中国超算行业有什么影响？
对中国正在研制中的三台E级超算影响不大
在2018年夏天，美国的超算SUMMIT终于排上了TOP500的第一名，把霸榜多年的中国“太湖之光”挤到了第二名，它比“太湖之光”快60%左右；接着，美国的另一台超算Sierra投入使用，它以微弱的优势把“太湖之光”挤到了第三位。
美国的媒体兴奋起来，一改多年的沉默，开始吹嘘下一代E级超算的威力。所谓E级超算，是指每秒钟运算一百亿亿次的超级计算机，比SUMMIT还好快许多倍。美国的E级超算计划在2021年左右投入使用，而中国为E级超算准备了三个方案，全部使用国产CPU和加速器，也计划在2020到2021年期间陆续投入使用。
这里需要介绍一下加速器，超算的核心部件不止是CPU，由于超算做的是密集的科学计算工作，CPU执行这样的计算任务并不是效率最高的，拥有众多运算单元的加速器处理器的效率更高，比较典型的加速器芯片就是Nvidia公司的通用型GPU，以及英特尔公司的至强Xeon Phi处理器。
最近几年出现的超算有一个趋势，就是使用一颗CPU带多颗加速器的架构，计算任务主要由加速器芯片完成，CPU主要做协调，通讯和分配任务的工作。
那么，美国的这次禁运对我们的E级超算有什么影响呢？让我们来仔细看看。
我国为下一代E级超算准备的三套方案，全都使用国产CPU和加速器芯片：
天河三号E级超算：
2018年5月19日，国家超算天津中心展示了天河三号原型机，在2018年底投入使用。天河三号的全配置主机计划在2020年投入使用，性能计划为1EFlops。 天河三号将使用国产飞腾FT2000plus或者后续型号的CPU，以及国防科技大学的Matrix2000plus或者后续型号的加速器芯片。
飞腾CPU采用ARM架构。
Matrix2000加速芯片是国防科大的通用DSP芯片，GPDSP，就像英伟达公司通用GPU（GPGPU）那样，DSP（数字信号处理器）也拥有大量的运算单元。
神威E级超算：
神威系列超算是由江南计算机研究所研制的。2018年7月神威E级原型机落户于国家超算济南中心，需要三年左右的研发时间，将在2021年正式投入使用。神威的CPU肯定是申威系列的最新型号，加速器可能也会采用Matrix2000plus或者其后续型号。
申威系列CPU脱胎于DEC Alpha处理器架构。
曙光E级超算：
中科曙光主要使用X86架构的CPU。在最新的TOP500榜单里，共有曙光公司制造的63台超算。除了一台以外，其它62台全部使用英特尔的至强CPU，有些超算还使用了英伟达的Tesla加速器芯片。
猛的一看，似乎这次禁运对曙光公司是致命的打击。因为英特尔和英伟达的芯片以后都买不到了。但是，曙光公司早就有了全国产化布局。秘密就在这62台之外的那一台超算上。
CPU：
这台型号为TC8600的超算是2018年投入使用的，它的CPU采用的是天津海光公司的Hygon Dhyana处理器，32核，2Ghz频率。天津海光公司在2016年初得到了美国AMD公司的第一代ZEN架构的永久授权，起点相当高，要知道AMD公司在2018年才推出第二代ZEN2的架构。海光3000系列CPU使用的是x86通用架构，完全支持Windows10 Pro， Windows Server2019， Linux Centos 7.x等主流操作系统，而且还有类似AMD EPYC的服务器版，最多32核64线程，基础频率2.0GHz，加速频率可达3.0GHz，而且涵盖了从4核到32核在内的多个产品线。
在这次禁运公布后，海光公司无法再得到AMD新的产品架构授权，但这并不影响已经得到的授权。海光公司已经组建了超过500人的研发团队，有能力在已经得到的授权的基础上，进一步开发新的产品，提升CPU的性能。这也意味着未来X86架构的CPU在中国可能会产生硬分叉，我们将从此与美国分道扬镳，走出自己的路来。
加速器芯片：
而这台曙光超算的加速器芯片更有意思：它的名称为Deep Computing Processor。根据公开的新闻报道，曙光的加速器芯片是基于我国寒武纪公司的AI处理器架构研制的，全球独一无二。曙光公司的通稿上说：“2017年11月，曙光基于寒武纪低功耗深度学习处理器，率先推出PHANERON，引起业界广泛关注。时隔6个月，寒武纪发布性能更为强劲的MLU100，曙光同步升级PHANERON产品，并发布一系列搭载寒武纪MLU100的服务器产品，保持着在智能服务器领域的持续领先”
这意味着曙光TC8600超算应用于AI领域将拥有极大的优势。
未来曙光公司的E级超算，应该就是基于海光公司的最新CPU和曙光自己设计的PHANERON AI加速芯片的最新版本。
由此也可以看出美国商务部的这次禁运，也根本影响不了曙光公司的E级超算产品。
禁运可能对曙光公司的商用超算产品线的影响巨大
我们还需要看到，由于曙光公司的主要产品线都是基于英特尔的CPU和英伟达的加速器芯片，在这次禁运之后，这些产品线都难以为继了。从商业层面来说，对曙光公司的影响巨大。但是好在曙光有了全国产的CPU和加速器芯片替代产品。我们可以从排名第43的曙光TC8600和排名第88位的曙光TC4600LP/W740l这两台超算的测试数据看出一些细节来：
表格来源于top500.org，笔者进行整理分析
曙光TC4600LP超算是基于英特尔至强CPU和英伟达Tesla加速器的产品。从以上测试数据来看，使用国产CPU和加速器的TC8600超算，在使用了3倍多的CPU核心数量以及5倍左右的加速器核心数量的情况下，跑分只不过增加到了1.4倍到1.7倍。但是功耗下降了很多，功耗效率提升了3倍多。
笔者分析这主要是因为TC8600采用的AI加速器并不适合传统的浮点运算，它只是在AI运算上具有巨大优势。
这也意味着，曙光公司缺乏合适的加速器芯片，来提升系统的浮点运算能力。这可能会对曙光公司的商业发展产生不利影响。
也许曙光公司的未来销售重点将转向AI运算。但是，也不排除未来曙光公司开发基于国产Matrix2000加速芯片的产品。
在这里必须介绍一下，为何曙光公司不能放弃X86架构的CPU。
全球TOP500强的超算100%都使用Linux操作系统的不同发行版。这些版本都为各自的CPU和加速器芯片做了优化。 而在超算上跑的应用软件也会受到CPU架构的影响。
从目前来看，对X86架构优化的Linux系统最成熟，应用软件最多，程序开发最容易，整个软件生态系统最成熟，稳定。而国产申威和飞腾CPU的系统则有些不足，软件工具为它们优化得少，编程也有些麻烦，系统bug也较多，整个软件生态系统只是解决了可用，离X86系统那样的好用仍然有一定距离。
所以，无论从国家的战略方面，还是从曙光公司的商业前景而言，X86架构都是应该坚持做下去的。
中国半导体制造技术的进步将极大地推进国产超算CPU和加速器芯片的性能
中国的半导体制造业长期以来是我们最大的短板。在台积电和三星电子已经量产7纳米工艺，英特尔已经小规模量产10纳米工艺的时候，我国目前最先进的半导体制造工艺还是中芯国际的28纳米工艺。
而在上面提到的国产CPU，要么使用国产较落后的工艺，自主性能够保证，但性能落后；要么在海外找半导体代工企业生产，性能可以提升，但自主性无法得到保证，说不定什么时候就会被掐脖子。
但这个状态很快就要结束了。中芯国际在2018年8月的第二季度业界报告中宣布，其14纳米生产工艺取得了重大进展，第一代FinFET技术研发已经进入了客户导入阶段。而正式量产则计划在2019年中期。所以，很快，我们就会拥有自己的14纳米半导体生产工艺了。
14纳米有多厉害？要知道，现在英特尔所有的台式机和服务器芯片都是14纳米的。英特尔在刚刚过去的台北电脑展上宣布会小规模量产10纳米工艺的低电压笔记本电脑处理器，并且预计未来几年14纳米仍然是它的主力工艺。只有AMD在台积电的代工下，刚刚宣布实现了CPU的7纳米工艺生产。
如果国产超算CPU和加速器芯片都使用中芯国际的14纳米工艺生产，它们的性能将会得到一个飞跃。
美国终将搬起石头砸自己的脚
实际上，美国在2015年对中国的四大超算中心的禁运，已经砸了它自己的脚：英特尔的至强Phi加速器芯片停产。
▲图为美国《个人计算机》杂志当年对美国封杀中国四大超算中心，禁止因特尔提供至强Phi加速器芯片给它们的报道
笔者在这里列一些事实：
中国的天河2号超算分别使用了3.2万颗Xeon E5 CPU和4.8万颗Xeon Phi加速器。
对于Xeon Phi的销量，Intel CEO曾表示过2016年的目标是10万颗出货量。
可见，单单我国的天河2号就能达到Intel 2016年约一半的Xeon Phi销量。
2015年4月9日，美国商务部宣布对中国的四家超算中心禁运英特尔的至强Xeon处理器和至强Phi加速器芯片。
2016年6月18日，至强Phi 7290/7250加速器价格直接腰斩。到2017年8月24日，英特尔宣布至强 Phi 7200系列协处理器加速卡进入停产退市进程。
▲图为美国《个人计算机》杂志在当年的报道中也提到中国市场对于因特尔这一芯片的重要性，以及美国禁令将给因特尔带来沉重打击
笔者毫不怀疑英特尔至强Phi加速器芯片的停产就是因为对中国的禁运。这甚至影响到了美国下一代E级超算的开发。美国的E级超算，CPU可选择英特尔和AMD的X86处理器以及IBM的Power处理器，而加速器芯片目前只有一个选择，就是英伟达的通用GPU。
那么美国商务部的这次禁运会有什么结果呢？笔者相信，在我们自力更生的努力之下，几年之后，世界上会出现几个新的超算CPU生态圈，彻底打破美国的垄断。
实际上，美国《纽约时报》和《华尔街日报》也都不认为美国政府的这一打击中国超算的举动会阻碍中国的发展。这两家媒体采访几位不同的美国超算专家就表示，美国政府的行为虽然短期内或给中国的超算发展带来阻碍，却也会更加坚定中国开发本土技术替换美国科技的决心。
而一件必然会要发生的事，就是在2020年和2021年，中国有三台E级超算将会冲击TOP500排行榜的冠军位置，把美国超算踢下去。让美国看看，谁才是超算领域的王者！
点击进入专题：
每日军情TOP5
The latest embargo in the United States will really make the Chinese next generation super fatal blow?
The latest embargo in the United States will really make the Chinese next generation super fatal blow?
588
In recent days, many American media have been schadenfreaked. They believe that China’s supercomputer industry will be hit hard, leaving the US CPU, and China will not make a supercomputer. Is this the truth? What is the impact of this ban on China's super-calculation industry?
According to several media reports, on June 21, local time, the US Department of Commerce once again listed five Chinese entities in the list of entities subject to import and export control on the grounds of national security, prohibiting them from purchasing parts from US suppliers. The relevant decision took effect on June 24. These five entities are related to two of the “Big Three” in China’s supercomputer industry, involving “Twilight” and “Shenwei”. The Jiangnan Institute of Computing Technology among the five entities is the R&D and manufacturing unit of Shenwei. The other four entities, Zhongke Shuguang, Tianjin Haiguang (Higon), Chengdu Haiguang Integrated Circuit, and Chengdu Haiguang Microelectronics Technology are all related to the “Dawn” series of supercomputers.
In recent days, many American media have been schadenfreaked. They believe that China’s supercomputer industry will be hit hard, leaving the US CPU, and China will not make a supercomputer. Is this the truth? What is the impact of this ban on China's super-calculation industry? This article will analyze this issue in detail.
First of all, why should the United States impose sanctions on China’s supercomputers again?
In fact, as early as the beginning of 2015, the US Department of Commerce listed China's four super-computing centers in the embargo list. This stems from the long-standing anxiety in the United States: in about 40 years, China's super-computing industry has completely imported American products, purchased CPUs from the United States to make super-computers, and then used its own CPU and accelerator to make super-calculations; A few super-calculations are not worth mentioning in the world's super-computing market. The number of super-calculated figures is the highest in the world, far surpassing that of the United States; from the low performance of the super-performance, it is not the highest, and the world's fastest super-calculation The first place in the rankings for many years; China's super-computing industry has come out of a perfect counterattack.
The world's supercomputer organization top500.org launched the world's fastest 500 supercomputer rankings every six months. The latest list has just been launched in June this year. The distribution map of the world's fastest 500 super-counting countries by country is as follows:
▲Data comes from top500.rog, which is organized by the author
In the TOP500 list, China's total number of supercomputers reached 219, accounting for 44% of the world's total, and the world's number one; the United States accounted for 23% of the world's 116 units, much behind China.
You know, in June 2018, just a year ago, China’s total number of supercomputers in the TOP500 was 206, and the United States was still 124. In this year, China has made progress and the United States is regressing. From the trend point of view, China's advantages will be greater and greater.
The anxiety in the American mind is understandable because of the enormous role of supercomputers in military, scientific research, and the economy. They were eager to curb China's development momentum, so they continued to enter the sewers with the principles of business integrity and fair trade, and banned the embargo. So, what impact does this embargo measure have on China's supercomputer industry?
It has little effect on the three E-class supercomputers under development in China.
In the summer of 2018, the United States' super-calculated SUMMIT finally ranked first in the TOP500, and pushed China's "Taihu Light", which has been in the dominance for many years, to the second place. It is about 60% faster than the "Taihu Light"; Another super-computer in the United States was put into use, and it pushed the "light of Taihu Lake" to the third place with a slight advantage.
The American media is excited, and after years of silence, it has begun to boast the power of the next generation of E-class supercomputer. The so-called E-class supercomputer refers to a supercomputer that operates 10 billion times per second, many times faster than SUMMIT. The US E-class super-calculation plan was put into use around 2021, and China has prepared three plans for E-class super-calculation, all using domestic CPUs and accelerators, and plans to use them from 2020 to 2021.
Here we need to introduce the accelerator. The core components of the supercomputer are not only the CPU. Because the supercomputer is doing intensive scientific computing work, the CPU does not perform the most efficient computing tasks. The accelerator processor with many computing units is more efficient. High, typical accelerator chips are Nvidia's general purpose GPUs and Intel's Xeon Peon processors.
A trend in the past few years has been the use of a CPU with multiple accelerator architectures. The computational tasks are mainly done by the accelerator chip. The CPU mainly performs coordination, communication and assignment tasks.
So, what impact does the US embargo have on our E-class supercomputer? Let's take a closer look.
China's three sets of solutions for the next-generation E-class supercomputer, all using domestic CPU and accelerator chips:
Tianhe No. 3 E-class super calculation:
On May 19, 2018, the National Supercomputer Tianjin Center demonstrated the Tianhe No. 3 prototype, which was put into use at the end of 2018. The fully-configured mainframe of Tianhe No. 3 is scheduled to be put into use in 2020 with a performance plan of 1EFlops. Tianhe No. 3 will use the domestic Feiteng FT2000plus or the subsequent models of the CPU, as well as the Defense Technology University's Matrix2000plus or subsequent models of accelerator chips.
The Feiteng CPU uses the ARM architecture.
The Matrix2000 acceleration chip is a general-purpose DSP chip of the National Defense Science and Technology University, GPDSP. Like the NVIDIA General GPU (GPGPU), the DSP (Digital Signal Processor) also has a large number of arithmetic units.
Shenwei E-class super calculation:
The Shenwei series supercomputer was developed by the Jiangnan Computer Research Institute. In July 2018, the Shenwei E-class prototype was settled in the National Supercomputer Jinan Center. It takes about three years of research and development time and will be officially put into use in 2021. Shenwei's CPU is definitely the latest model of the Shenwei series, and the accelerator may also use Matrix2000plus or its subsequent models.
The Shenwei series CPU was born out of the DEC Alpha processor architecture.
Dawning E-class super-calculation:
Zhongke Shuguang mainly uses CPUs of X86 architecture. In the latest TOP500 list, there are 63 supercomputers made by Shuguang. In addition to one, the other 62 all use Intel's Xeon CPU, and some supercomputers also use NVIDIA's Tesla accelerator chip.
A fierce look, it seems that the embargo is a fatal blow to the company. Because Intel and NVIDIA chips can't be bought in the future. However, Dawning has long had a national production layout. The secret is on the supercomputer outside of the 62 units.
CPU:
The super model of the TC8600 is put into use in 2018. Its CPU uses the Hygon Dhyana processor of Tianjin Haiguang Company, 32 cores, 2Ghz frequency. At the beginning of 2016, Tianjin Haiguang Company obtained the permanent authorization of the first generation ZEN architecture of American AMD Company. The starting point is quite high. It is necessary to know that AMD will launch the second generation ZEN2 architecture in 2018. The Haiguang 3000 series CPU uses the x86 universal architecture, which fully supports Windows10 Pro, Windows Server 2019, Linux Centos 7.x and other mainstream operating systems, and also has a server version similar to AMD EPYC, with a maximum of 32 cores and 64 threads, and a base frequency of 2.0 GHz. The acceleration frequency is up to 3.0GHz and covers multiple product lines from 4 cores to 32 cores.
After the embargo was announced, Haiguang could no longer obtain AMD's new product architecture license, but this did not affect the authorization already obtained. Haiguang has set up a research and development team of more than 500 people, and has the ability to further develop new products and improve CPU performance based on the already obtained authorization. This also means that the CPU of the X86 architecture may have a hard fork in China in the future, and we will part ways from the United States and go out of our way.
Accelerator chip:
And this Dawning Super Accelerator chip is more interesting: its name is Deep Computing Processor. According to public news reports, Dawning's accelerator chip is based on the AI processor architecture of China's Cambrian company, which is unique in the world. Dawning's manuscript said: "In November 2017, Dawning was the first to introduce PHANERON based on the Cambrian low-power deep learning processor, which attracted widespread attention in the industry. After 6 months, the Cambrian release performance was stronger. MLU100, Dawning upgrades PHANERON products simultaneously, and launches a series of server products equipped with Cambrian MLU100, maintaining a leading position in the field of intelligent servers"
This means that the Dawning TC8600 supercomputer will have great advantages in the AI field.
In the future, Dawning's E-class supercomputer should be based on the latest version of Haiguang's latest CPU and Dawning's own PHANERON AI acceleration chip.
It can also be seen that the embargo of the US Department of Commerce does not affect the company's E-class super-computing products.
The embargo may have a huge impact on Dawning's commercial super-computing product line
We also need to see that because Dawning's main product line is based on Intel's CPU and NVIDIA's accelerator chip, after the embargo, these product lines are unsustainable. From a commercial perspective, the impact on the company is huge. But fortunately, Dawn has a nationally produced CPU and accelerator chip replacement. We can see some details from the super-calculated test data of the 43rd Dawning TC8600 and the 88th Dawning TC4600LP/W740l:
The form is from top500.org, and the author analyzes it.
The Dawning TC4600LP Supercomputer is based on the Intel Xeon CPU and the NVIDIA Tesla Accelerator. From the above test data, the TC8600 supercomputer using the domestic CPU and accelerator, in the case of using more than 3 times the number of CPU cores and 5 times the number of accelerator cores, the running points only increased by 1.4 times to 1.7 times. . However, the power consumption has dropped a lot, and the power efficiency has increased by more than three times.
The author analyzes this mainly because the AI accelerator used by the TC8600 is not suitable for traditional floating-point operations, it only has great advantages in AI computing.
This also means that Dawning lacks a suitable accelerator chip to improve the system's floating-point computing power. This may adversely affect Dawning's business development.
Perhaps Dawning's future sales focus will shift to AI computing. However, it is not ruled out that Dawning will develop products based on the domestic Matrix2000 acceleration chip.
Here we must introduce why the company can not give up the CPU of the X86 architecture.
100% of the world's top TOP500 supercomputers use different distributions of the Linux operating system. These versions are optimized for their respective CPU and accelerator chips. Applications running on supercomputers are also affected by the CPU architecture.
From the current point of view, the Linux system optimized for the X86 architecture is the most mature, the application software is the most, the program development is the easiest, and the entire software ecosystem is the most mature and stable. The domestic Shenwei and Feiteng CPU systems are somewhat inadequate. The software tools are less optimized for them, the programming is also somewhat troublesome, and the system bugs are more. The entire software ecosystem is only available, and it is still useful like the X86 system. a certain distance.
Therefore, regardless of the strategic aspect of the country or the business prospects of the company, the X86 architecture should continue to do so.
Advances in China's semiconductor manufacturing technology will greatly advance the performance of domestic supercomputer and accelerator chips
China's semiconductor manufacturing industry has long been our biggest shortcoming. At the time when TSMC and Samsung Electronics have mass-produced the 7-nanometer process, Intel has already produced a small-scale mass production of 10 nanometers. At present, China's most advanced semiconductor manufacturing process is still the 28-nanometer process of SMIC.
In the domestic CPU mentioned above, either the domestically-developed process is used, the autonomy can be guaranteed, but the performance is backward; or the semiconductor foundry is found overseas, the performance can be improved, but the autonomy cannot be guaranteed, maybe I will be licked at the neck.
But this state is coming to an end soon. In the second quarter of August 2018, SMIC announced that its 14nm production process has made significant progress, and the first generation of FinFET technology research and development has entered the customer introduction stage. The formal mass production is planned for the middle of 2019. So, soon, we will have our own 14nm semiconductor manufacturing process.
How powerful is 14 nanometers? You know, all of Intel's desktop and server chips are now 14 nanometers. Intel announced at the Taipei Computer Show that it will produce a low-voltage notebook processor with a small-scale production of 10 nanometers. It is expected that 14 nanometers will remain its main process in the next few years. Only AMD, under the OEM of TSMC, just announced the realization of the 7-nanometer process of CPU production.
If the domestic supercomputer and accelerator chips are produced using SMIC's 14-nanometer process, their performance will be a leap.
The United States will eventually lift a rock and lick its own feet.
In fact, the US embargo on China's four supercomputer centers in 2015 has already smashed its own feet: Intel's Xeon Phi accelerator chip was discontinued.
▲The picture shows that the US "Personal Computer" magazine blocked the four super-computing centers in China for the United States, and banned Intel's coverage of the Xeon Phi accelerator chip.
The author lists some facts here:
China's Tianhe No. 2 supercomputer uses 32,000 Xeon E5 CPUs and 48,000 Xeon Phi accelerators respectively.
For Xeon Phi sales, Intel CEO has stated that the target for 2016 is 100,000 shipments.
It can be seen that China's Tianhe 2 alone can reach about half of Intel's Xeon Phi sales in 2016.
On April 9, 2015, the US Department of Commerce announced that it has banned Intel's Xeon Xeon processor and Xeon Phi accelerator chip for four Chinese supercomputer centers.
On June 18, 2016, the Xeon Phi 7290/7250 accelerator price was directly at the waist. On August 24, 2017, Intel announced that the Xeon Phi 7200 series coprocessor accelerator card entered the discontinuation process.
▲The picture shows that the US "Personal Computer" magazine also mentioned the Chinese market for Inter in the report of the year.
美国的最新禁运 真的会让中国下代超算遭致命打击？
美国的最新禁运 真的会让中国下代超算遭致命打击？
588
最近这几天，有许多美国媒体幸灾乐祸，认为中国超算行业将受到致命打击，离开了美国的CPU，中国将造不出超算。事实是这样的吗？这个禁令对我国的超算行业有什么影响？
据多家媒体报道，在当地时间6月21日，美国商务部以国家安全为由，再次将中国的5家实体列入进出口管制的实体清单，禁止它们从美国的供应商采购零部件，相关决定于6月24日生效。这五家实体与中国超算行业的“三巨头”中的两位相关，涉及到“曙光”和“神威”两款超算。这五家实体中的江南计算技术研究所就是“神威”超算的研发制造单位。而其他四家实体，中科曙光，天津海光（Higon）、成都海光集成电路、成都海光微电子技术全都与“曙光”系列超算有关。
最近这几天，有许多美国媒体幸灾乐祸，认为中国超算行业将受到致命打击，离开了美国的CPU，中国将造不出超算。事实是这样的吗？这个禁令对我国的超算行业有什么影响？本文将对这个问题进行详细分析。
首先，美国为什么要再次对我国超算企业进行制裁？
实际上，早在2015年初，美国商务部就将中国的四大超算中心列入了禁运清单。这源于美国长久以来的焦虑：在大约40年中，中国的超算行业从完全进口美国产品，到购买美国的CPU自己制造超算，再到使用自己的CPU和加速器制造超算； 从只有几台超算，在全世界超算市场里不值一提，到拥有超算的数量全球第一，大大超越美国；从拥有的超算性能低下排不上号，到称霸全球最快超算排行榜第一名多年；中国超算行业走出了一条完美的逆袭之路。
世界超算组织top500.org每半年推出一次全球最快的500台超算排行榜。今年6月刚刚推出了最新的榜单。全球最快的500台超算数量按国家的分布图如下：
▲数据来源于top500.rog，由笔者整理
TOP500榜单中，中国超算的数量达到了219台，占全球的44%，世界第一；美国以116台的数量占全球的23%，比中国落后甚多。
要知道，在仅仅一年前的2018年6月，中国在TOP500中的超算总数还是206台，美国还是124台。这一年中，中国又有了进步，美国又在退步。从趋势来看，中国的优势会越来越大。
由于超算在军事，科研和经济中的巨大作用，美国心里的焦虑是可以理解的。他们急于抑制住中国的发展势头，于是就把商业诚信和公平贸易原则仍进下水道，祭出了禁运大法。那么，这次的禁运措施对中国超算行业有什么影响？
对中国正在研制中的三台E级超算影响不大
在2018年夏天，美国的超算SUMMIT终于排上了TOP500的第一名，把霸榜多年的中国“太湖之光”挤到了第二名，它比“太湖之光”快60%左右；接着，美国的另一台超算Sierra投入使用，它以微弱的优势把“太湖之光”挤到了第三位。
美国的媒体兴奋起来，一改多年的沉默，开始吹嘘下一代E级超算的威力。所谓E级超算，是指每秒钟运算一百亿亿次的超级计算机，比SUMMIT还好快许多倍。美国的E级超算计划在2021年左右投入使用，而中国为E级超算准备了三个方案，全部使用国产CPU和加速器，也计划在2020到2021年期间陆续投入使用。
这里需要介绍一下加速器，超算的核心部件不止是CPU，由于超算做的是密集的科学计算工作，CPU执行这样的计算任务并不是效率最高的，拥有众多运算单元的加速器处理器的效率更高，比较典型的加速器芯片就是Nvidia公司的通用型GPU，以及英特尔公司的至强Xeon Phi处理器。
最近几年出现的超算有一个趋势，就是使用一颗CPU带多颗加速器的架构，计算任务主要由加速器芯片完成，CPU主要做协调，通讯和分配任务的工作。
那么，美国的这次禁运对我们的E级超算有什么影响呢？让我们来仔细看看。
我国为下一代E级超算准备的三套方案，全都使用国产CPU和加速器芯片：
天河三号E级超算：
2018年5月19日，国家超算天津中心展示了天河三号原型机，在2018年底投入使用。天河三号的全配置主机计划在2020年投入使用，性能计划为1EFlops。 天河三号将使用国产飞腾FT2000plus或者后续型号的CPU，以及国防科技大学的Matrix2000plus或者后续型号的加速器芯片。
飞腾CPU采用ARM架构。
Matrix2000加速芯片是国防科大的通用DSP芯片，GPDSP，就像英伟达公司通用GPU（GPGPU）那样，DSP（数字信号处理器）也拥有大量的运算单元。
神威E级超算：
神威系列超算是由江南计算机研究所研制的。2018年7月神威E级原型机落户于国家超算济南中心，需要三年左右的研发时间，将在2021年正式投入使用。神威的CPU肯定是申威系列的最新型号，加速器可能也会采用Matrix2000plus或者其后续型号。
申威系列CPU脱胎于DEC Alpha处理器架构。
曙光E级超算：
中科曙光主要使用X86架构的CPU。在最新的TOP500榜单里，共有曙光公司制造的63台超算。除了一台以外，其它62台全部使用英特尔的至强CPU，有些超算还使用了英伟达的Tesla加速器芯片。
猛的一看，似乎这次禁运对曙光公司是致命的打击。因为英特尔和英伟达的芯片以后都买不到了。但是，曙光公司早就有了全国产化布局。秘密就在这62台之外的那一台超算上。
CPU：
这台型号为TC8600的超算是2018年投入使用的，它的CPU采用的是天津海光公司的Hygon Dhyana处理器，32核，2Ghz频率。天津海光公司在2016年初得到了美国AMD公司的第一代ZEN架构的永久授权，起点相当高，要知道AMD公司在2018年才推出第二代ZEN2的架构。海光3000系列CPU使用的是x86通用架构，完全支持Windows10 Pro， Windows Server2019， Linux Centos 7.x等主流操作系统，而且还有类似AMD EPYC的服务器版，最多32核64线程，基础频率2.0GHz，加速频率可达3.0GHz，而且涵盖了从4核到32核在内的多个产品线。
在这次禁运公布后，海光公司无法再得到AMD新的产品架构授权，但这并不影响已经得到的授权。海光公司已经组建了超过500人的研发团队，有能力在已经得到的授权的基础上，进一步开发新的产品，提升CPU的性能。这也意味着未来X86架构的CPU在中国可能会产生硬分叉，我们将从此与美国分道扬镳，走出自己的路来。
加速器芯片：
而这台曙光超算的加速器芯片更有意思：它的名称为Deep Computing Processor。根据公开的新闻报道，曙光的加速器芯片是基于我国寒武纪公司的AI处理器架构研制的，全球独一无二。曙光公司的通稿上说：“2017年11月，曙光基于寒武纪低功耗深度学习处理器，率先推出PHANERON，引起业界广泛关注。时隔6个月，寒武纪发布性能更为强劲的MLU100，曙光同步升级PHANERON产品，并发布一系列搭载寒武纪MLU100的服务器产品，保持着在智能服务器领域的持续领先”
这意味着曙光TC8600超算应用于AI领域将拥有极大的优势。
未来曙光公司的E级超算，应该就是基于海光公司的最新CPU和曙光自己设计的PHANERON AI加速芯片的最新版本。
由此也可以看出美国商务部的这次禁运，也根本影响不了曙光公司的E级超算产品。
禁运可能对曙光公司的商用超算产品线的影响巨大
我们还需要看到，由于曙光公司的主要产品线都是基于英特尔的CPU和英伟达的加速器芯片，在这次禁运之后，这些产品线都难以为继了。从商业层面来说，对曙光公司的影响巨大。但是好在曙光有了全国产的CPU和加速器芯片替代产品。我们可以从排名第43的曙光TC8600和排名第88位的曙光TC4600LP/W740l这两台超算的测试数据看出一些细节来：
表格来源于top500.org，笔者进行整理分析
曙光TC4600LP超算是基于英特尔至强CPU和英伟达Tesla加速器的产品。从以上测试数据来看，使用国产CPU和加速器的TC8600超算，在使用了3倍多的CPU核心数量以及5倍左右的加速器核心数量的情况下，跑分只不过增加到了1.4倍到1.7倍。但是功耗下降了很多，功耗效率提升了3倍多。
笔者分析这主要是因为TC8600采用的AI加速器并不适合传统的浮点运算，它只是在AI运算上具有巨大优势。
这也意味着，曙光公司缺乏合适的加速器芯片，来提升系统的浮点运算能力。这可能会对曙光公司的商业发展产生不利影响。
也许曙光公司的未来销售重点将转向AI运算。但是，也不排除未来曙光公司开发基于国产Matrix2000加速芯片的产品。
在这里必须介绍一下，为何曙光公司不能放弃X86架构的CPU。
全球TOP500强的超算100%都使用Linux操作系统的不同发行版。这些版本都为各自的CPU和加速器芯片做了优化。 而在超算上跑的应用软件也会受到CPU架构的影响。
从目前来看，对X86架构优化的Linux系统最成熟，应用软件最多，程序开发最容易，整个软件生态系统最成熟，稳定。而国产申威和飞腾CPU的系统则有些不足，软件工具为它们优化得少，编程也有些麻烦，系统bug也较多，整个软件生态系统只是解决了可用，离X86系统那样的好用仍然有一定距离。
所以，无论从国家的战略方面，还是从曙光公司的商业前景而言，X86架构都是应该坚持做下去的。
中国半导体制造技术的进步将极大地推进国产超算CPU和加速器芯片的性能
中国的半导体制造业长期以来是我们最大的短板。在台积电和三星电子已经量产7纳米工艺，英特尔已经小规模量产10纳米工艺的时候，我国目前最先进的半导体制造工艺还是中芯国际的28纳米工艺。
而在上面提到的国产CPU，要么使用国产较落后的工艺，自主性能够保证，但性能落后；要么在海外找半导体代工企业生产，性能可以提升，但自主性无法得到保证，说不定什么时候就会被掐脖子。
但这个状态很快就要结束了。中芯国际在2018年8月的第二季度业界报告中宣布，其14纳米生产工艺取得了重大进展，第一代FinFET技术研发已经进入了客户导入阶段。而正式量产则计划在2019年中期。所以，很快，我们就会拥有自己的14纳米半导体生产工艺了。
14纳米有多厉害？要知道，现在英特尔所有的台式机和服务器芯片都是14纳米的。英特尔在刚刚过去的台北电脑展上宣布会小规模量产10纳米工艺的低电压笔记本电脑处理器，并且预计未来几年14纳米仍然是它的主力工艺。只有AMD在台积电的代工下，刚刚宣布实现了CPU的7纳米工艺生产。
如果国产超算CPU和加速器芯片都使用中芯国际的14纳米工艺生产，它们的性能将会得到一个飞跃。
美国终将搬起石头砸自己的脚
实际上，美国在2015年对中国的四大超算中心的禁运，已经砸了它自己的脚：英特尔的至强Phi加速器芯片停产。
▲图为美国《个人计算机》杂志当年对美国封杀中国四大超算中心，禁止因特尔提供至强Phi加速器芯片给它们的报道
笔者在这里列一些事实：
中国的天河2号超算分别使用了3.2万颗Xeon E5 CPU和4.8万颗Xeon Phi加速器。
对于Xeon Phi的销量，Intel CEO曾表示过2016年的目标是10万颗出货量。
可见，单单我国的天河2号就能达到Intel 2016年约一半的Xeon Phi销量。
2015年4月9日，美国商务部宣布对中国的四家超算中心禁运英特尔的至强Xeon处理器和至强Phi加速器芯片。
2016年6月18日，至强Phi 7290/7250加速器价格直接腰斩。到2017年8月24日，英特尔宣布至强 Phi 7200系列协处理器加速卡进入停产退市进程。
▲图为美国《个人计算机》杂志在当年的报道中也提到中国市场对于因特尔这一芯片的重要性，以及美国禁令将给因特尔带来沉重打击
笔者毫不怀疑英特尔至强Phi加速器芯片的停产就是因为对中国的禁运。这甚至影响到了美国下一代E级超算的开发。美国的E级超算，CPU可选择英特尔和AMD的X86处理器以及IBM的Power处理器，而加速器芯片目前只有一个选择，就是英伟达的通用GPU。
那么美国商务部的这次禁运会有什么结果呢？笔者相信，在我们自力更生的努力之下，几年之后，世界上会出现几个新的超算CPU生态圈，彻底打破美国的垄断。
实际上，美国《纽约时报》和《华尔街日报》也都不认为美国政府的这一打击中国超算的举动会阻碍中国的发展。这两家媒体采访几位不同的美国超算专家就表示，美国政府的行为虽然短期内或给中国的超算发展带来阻碍，却也会更加坚定中国开发本土技术替换美国科技的决心。
而一件必然会要发生的事，就是在2020年和2021年，中国有三台E级超算将会冲击TOP500排行榜的冠军位置，把美国超算踢下去。让美国看看，谁才是超算领域的王者！
点击进入专题：
每日军情TOP5
The latest embargo in the United States will really make the Chinese next generation super fatal blow?
The latest embargo in the United States will really make the Chinese next generation super fatal blow?
588
In recent days, many American media have been schadenfreaked. They believe that China’s supercomputer industry will be hit hard, leaving the US CPU, and China will not make a supercomputer. Is this the truth? What is the impact of this ban on China's super-calculation industry?
According to several media reports, on June 21, local time, the US Department of Commerce once again listed five Chinese entities in the list of entities subject to import and export control on the grounds of national security, prohibiting them from purchasing parts from US suppliers. The relevant decision took effect on June 24. These five entities are related to two of the “Big Three” in China’s supercomputer industry, involving “Twilight” and “Shenwei”. The Jiangnan Institute of Computing Technology among the five entities is the R&D and manufacturing unit of Shenwei. The other four entities, Zhongke Shuguang, Tianjin Haiguang (Higon), Chengdu Haiguang Integrated Circuit, and Chengdu Haiguang Microelectronics Technology are all related to the “Dawn” series of supercomputers.
In recent days, many American media have been schadenfreaked. They believe that China’s supercomputer industry will be hit hard, leaving the US CPU, and China will not make a supercomputer. Is this the truth? What is the impact of this ban on China's super-calculation industry? This article will analyze this issue in detail.
First of all, why should the United States impose sanctions on China’s supercomputers again?
In fact, as early as the beginning of 2015, the US Department of Commerce listed China's four super-computing centers in the embargo list. This stems from the long-standing anxiety in the United States: in about 40 years, China's super-computing industry has completely imported American products, purchased CPUs from the United States to make super-computers, and then used its own CPU and accelerator to make super-calculations; A few super-calculations are not worth mentioning in the world's super-computing market. The number of super-calculated figures is the highest in the world, far surpassing that of the United States; from the low performance of the super-performance, it is not the highest, and the world's fastest super-calculation The first place in the rankings for many years; China's super-computing industry has come out of a perfect counterattack.
The world's supercomputer organization top500.org launched the world's fastest 500 supercomputer rankings every six months. The latest list has just been launched in June this year. The distribution map of the world's fastest 500 super-counting countries by country is as follows:
▲Data comes from top500.rog, which is organized by the author
In the TOP500 list, China's total number of supercomputers reached 219, accounting for 44% of the world's total, and the world's number one; the United States accounted for 23% of the world's 116 units, much behind China.
You know, in June 2018, just a year ago, China’s total number of supercomputers in the TOP500 was 206, and the United States was still 124. In this year, China has made progress and the United States is regressing. From the trend point of view, China's advantages will be greater and greater.
The anxiety in the American mind is understandable because of the enormous role of supercomputers in military, scientific research, and the economy. They were eager to curb China's development momentum, so they continued to enter the sewers with the principles of business integrity and fair trade, and banned the embargo. So, what impact does this embargo measure have on China's supercomputer industry?
It has little effect on the three E-class supercomputers under development in China.
In the summer of 2018, the United States' super-calculated SUMMIT finally ranked first in the TOP500, and pushed China's "Taihu Light", which has been in the dominance for many years, to the second place. It is about 60% faster than the "Taihu Light"; Another super-computer in the United States was put into use, and it pushed the "light of Taihu Lake" to the third place with a slight advantage.
The American media is excited, and after years of silence, it has begun to boast the power of the next generation of E-class supercomputer. The so-called E-class supercomputer refers to a supercomputer that operates 10 billion times per second, many times faster than SUMMIT. The US E-class super-calculation plan was put into use around 2021, and China has prepared three plans for E-class super-calculation, all using domestic CPUs and accelerators, and plans to use them from 2020 to 2021.
Here we need to introduce the accelerator. The core components of the supercomputer are not only the CPU. Because the supercomputer is doing intensive scientific computing work, the CPU does not perform the most efficient computing tasks. The accelerator processor with many computing units is more efficient. High, typical accelerator chips are Nvidia's general purpose GPUs and Intel's Xeon Peon processors.
A trend in the past few years has been the use of a CPU with multiple accelerator architectures. The computational tasks are mainly done by the accelerator chip. The CPU mainly performs coordination, communication and assignment tasks.
So, what impact does the US embargo have on our E-class supercomputer? Let's take a closer look.
China's three sets of solutions for the next-generation E-class supercomputer, all using domestic CPU and accelerator chips:
Tianhe No. 3 E-class super calculation:
On May 19, 2018, the National Supercomputer Tianjin Center demonstrated the Tianhe No. 3 prototype, which was put into use at the end of 2018. The fully-configured mainframe of Tianhe No. 3 is scheduled to be put into use in 2020 with a performance plan of 1EFlops. Tianhe No. 3 will use the domestic Feiteng FT2000plus or the subsequent models of the CPU, as well as the Defense Technology University's Matrix2000plus or subsequent models of accelerator chips.
The Feiteng CPU uses the ARM architecture.
The Matrix2000 acceleration chip is a general-purpose DSP chip of the National Defense Science and Technology University, GPDSP. Like the NVIDIA General GPU (GPGPU), the DSP (Digital Signal Processor) also has a large number of arithmetic units.
Shenwei E-class super calculation:
The Shenwei series supercomputer was developed by the Jiangnan Computer Research Institute. In July 2018, the Shenwei E-class prototype was settled in the National Supercomputer Jinan Center. It takes about three years of research and development time and will be officially put into use in 2021. Shenwei's CPU is definitely the latest model of the Shenwei series, and the accelerator may also use Matrix2000plus or its subsequent models.
The Shenwei series CPU was born out of the DEC Alpha processor architecture.
Dawning E-class super-calculation:
Zhongke Shuguang mainly uses CPUs of X86 architecture. In the latest TOP500 list, there are 63 supercomputers made by Shuguang. In addition to one, the other 62 all use Intel's Xeon CPU, and some supercomputers also use NVIDIA's Tesla accelerator chip.
A fierce look, it seems that the embargo is a fatal blow to the company. Because Intel and NVIDIA chips can't be bought in the future. However, Dawning has long had a national production layout. The secret is on the supercomputer outside of the 62 units.
CPU:
The super model of the TC8600 is put into use in 2018. Its CPU uses the Hygon Dhyana processor of Tianjin Haiguang Company, 32 cores, 2Ghz frequency. At the beginning of 2016, Tianjin Haiguang Company obtained the permanent authorization of the first generation ZEN architecture of American AMD Company. The starting point is quite high. It is necessary to know that AMD will launch the second generation ZEN2 architecture in 2018. The Haiguang 3000 series CPU uses the x86 universal architecture, which fully supports Windows10 Pro, Windows Server 2019, Linux Centos 7.x and other mainstream operating systems, and also has a server version similar to AMD EPYC, with a maximum of 32 cores and 64 threads, and a base frequency of 2.0 GHz. The acceleration frequency is up to 3.0GHz and covers multiple product lines from 4 cores to 32 cores.
After the embargo was announced, Haiguang could no longer obtain AMD's new product architecture license, but this did not affect the authorization already obtained. Haiguang has set up a research and development team of more than 500 people, and has the ability to further develop new products and improve CPU performance based on the already obtained authorization. This also means that the CPU of the X86 architecture may have a hard fork in China in the future, and we will part ways from the United States and go out of our way.
Accelerator chip:
And this Dawning Super Accelerator chip is more interesting: its name is Deep Computing Processor. According to public news reports, Dawning's accelerator chip is based on the AI processor architecture of China's Cambrian company, which is unique in the world. Dawning's manuscript said: "In November 2017, Dawning was the first to introduce PHANERON based on the Cambrian low-power deep learning processor, which attracted widespread attention in the industry. After 6 months, the Cambrian release performance was stronger. MLU100, Dawning upgrades PHANERON products simultaneously, and launches a series of server products equipped with Cambrian MLU100, maintaining a leading position in the field of intelligent servers"
This means that the Dawning TC8600 supercomputer will have great advantages in the AI field.
In the future, Dawning's E-class supercomputer should be based on the latest version of Haiguang's latest CPU and Dawning's own PHANERON AI acceleration chip.
It can also be seen that the embargo of the US Department of Commerce does not affect the company's E-class super-computing products.
The embargo may have a huge impact on Dawning's commercial super-computing product line
We also need to see that because Dawning's main product line is based on Intel's CPU and NVIDIA's accelerator chip, after the embargo, these product lines are unsustainable. From a commercial perspective, the impact on the company is huge. But fortunately, Dawn has a nationally produced CPU and accelerator chip replacement. We can see some details from the super-calculated test data of the 43rd Dawning TC8600 and the 88th Dawning TC4600LP/W740l:
The form is from top500.org, and the author analyzes it.
The Dawning TC4600LP Supercomputer is based on the Intel Xeon CPU and the NVIDIA Tesla Accelerator. From the above test data, the TC8600 supercomputer using the domestic CPU and accelerator, in the case of using more than 3 times the number of CPU cores and 5 times the number of accelerator cores, the running points only increased by 1.4 times to 1.7 times. . However, the power consumption has dropped a lot, and the power efficiency has increased by more than three times.
The author analyzes this mainly because the AI accelerator used by the TC8600 is not suitable for traditional floating-point operations, it only has great advantages in AI computing.
This also means that Dawning lacks a suitable accelerator chip to improve the system's floating-point computing power. This may adversely affect Dawning's business development.
Perhaps Dawning's future sales focus will shift to AI computing. However, it is not ruled out that Dawning will develop products based on the domestic Matrix2000 acceleration chip.
Here we must introduce why the company can not give up the CPU of the X86 architecture.
100% of the world's top TOP500 supercomputers use different distributions of the Linux operating system. These versions are optimized for their respective CPU and accelerator chips. Applications running on supercomputers are also affected by the CPU architecture.
From the current point of view, the Linux system optimized for the X86 architecture is the most mature, the application software is the most, the program development is the easiest, and the entire software ecosystem is the most mature and stable. The domestic Shenwei and Feiteng CPU systems are somewhat inadequate. The software tools are less optimized for them, the programming is also somewhat troublesome, and the system bugs are more. The entire software ecosystem is only available, and it is still useful like the X86 system. a certain distance.
Therefore, regardless of the strategic aspect of the country or the business prospects of the company, the X86 architecture should continue to do so.
Advances in China's semiconductor manufacturing technology will greatly advance the performance of domestic supercomputer and accelerator chips
China's semiconductor manufacturing industry has long been our biggest shortcoming. At the time when TSMC and Samsung Electronics have mass-produced the 7-nanometer process, Intel has already produced a small-scale mass production of 10 nanometers. At present, China's most advanced semiconductor manufacturing process is still the 28-nanometer process of SMIC.
In the domestic CPU mentioned above, either the domestically-developed process is used, the autonomy can be guaranteed, but the performance is backward; or the semiconductor foundry is found overseas, the performance can be improved, but the autonomy cannot be guaranteed, maybe I will be licked at the neck.
But this state is coming to an end soon. In the second quarter of August 2018, SMIC announced that its 14nm production process has made significant progress, and the first generation of FinFET technology research and development has entered the customer introduction stage. The formal mass production is planned for the middle of 2019. So, soon, we will have our own 14nm semiconductor manufacturing process.
How powerful is 14 nanometers? You know, all of Intel's desktop and server chips are now 14 nanometers. Intel announced at the Taipei Computer Show that it will produce a low-voltage notebook processor with a small-scale production of 10 nanometers. It is expected that 14 nanometers will remain its main process in the next few years. Only AMD, under the OEM of TSMC, just announced the realization of the 7-nanometer process of CPU production.
If the domestic supercomputer and accelerator chips are produced using SMIC's 14-nanometer process, their performance will be a leap.
The United States will eventually lift a rock and lick its own feet.
In fact, the US embargo on China's four supercomputer centers in 2015 has already smashed its own feet: Intel's Xeon Phi accelerator chip was discontinued.
▲The picture shows that the US "Personal Computer" magazine blocked the four super-computing centers in China for the United States, and banned Intel's coverage of the Xeon Phi accelerator chip.
The author lists some facts here:
China's Tianhe No. 2 supercomputer uses 32,000 Xeon E5 CPUs and 48,000 Xeon Phi accelerators respectively.
For Xeon Phi sales, Intel CEO has stated that the target for 2016 is 100,000 shipments.
It can be seen that China's Tianhe 2 alone can reach about half of Intel's Xeon Phi sales in 2016.
On April 9, 2015, the US Department of Commerce announced that it has banned Intel's Xeon Xeon processor and Xeon Phi accelerator chip for four Chinese supercomputer centers.
On June 18, 2016, the Xeon Phi 7290/7250 accelerator price was directly at the waist. On August 24, 2017, Intel announced that the Xeon Phi 7200 series coprocessor accelerator card entered the discontinuation process.
▲The picture shows that the US "Personal Computer" magazine also mentioned the Chinese market for Inter in the report of the year.
