• Monday, May 27, 2019

More than AI, Huawei exposed the world's first 7nm Arm server chip

Discussion in 'China & Far East' started by Adam WANG SHANGHAI MEGA, Oct 11, 2018.



    Aug 7, 2017
    +2 / 9,099 / -17
    More than AI, Huawei exposed the world's first 7nm Arm server chip
    2018-10-11 10:39:12

    At the Huawei Connected Conference held yesterday, Huawei finally unveiled the mystery of its self-developed AI chip.


    According to Xu Zhijun, president of Huawei's rotation, the two chips that Huawei launched this time are the Shengteng 910 and the Shengteng 310. These are two new products that Huawei created based on its self-developed DaVinci architecture. Among them, the Shengteng 910 is the AI chip with the highest calculation density of the current single chip. The product is built in 7nm process, the maximum power consumption is 350w, and other parameters are also superior: 256 TeraFLOPS can be achieved under half precision (FP16), and 512 TeraOPS is achieved under integer precision (INT8). The chip also supports a 128-channel full HD video decoder (H.264/265).

    Huawei Shengteng 910 Introduction

    Xu Zhijun said that Huawei's Shengteng 910 also has a strong lead when compared with NVIDIA and Google's chips. The chip will arrive in Q2 in 2019, which will give Huawei a strong support for the training and logic service series in the cloud, breaking through the current market monopoly by TPU and Nvidia.

    Huawei Shengteng 910 Introduction

    In addition, Huawei also released the Shengteng 310. According to Xu Zhijun, Huawei's extremely efficient and low-power AI SoC is launched for edge AI. As a chip that also uses the DaVinci architecture, the Huawei Shengteng 310 is manufactured using a 12nm FFC process. In the case of half-precision (FP16), it can achieve 8 TeraFLOPS. Under integer precision (INT8), it can be done. 16 TeraOPS, which also supports 16-channel full HD video decoder - H.264/265, and its maximum power consumption is only 8W, this chip is now able to provide customers with full support.

    Huawei Shengteng 310 Introduction

    Huawei said that their series of AI IP and chip upgrades based on a unified and scalable architecture have five series of nano, tiny, mini, lite and max, which can provide optimal TOPS/W support across the entire scene.

    In an interview, Xu Zhijun emphasized to reporters that Huawei's Shengteng chip will not be sold separately, but will be sold in the form of AI accelerator card, acceleration module, server and all-in-one. Huawei's full-stack AI strategy has also been fully complemented after its launch.

    Huawei's AI solution

    In the future-oriented AI opportunities, Huawei will focus on investing in basic research, building a full-stack solution, investing in open ecology and talent development, solution enhancement and internal efficiency improvement. Specifically:

    Constructing data efficient (less data requirements), energy efficient (lower computing power and energy consumption), safe and credible, auto-autonomous machine learning fundamentals in areas such as computational vision, natural language processing, and decision-making reasoning;
    Create a complete, collaborative, and full-stack solution for cloud, edge and end-to-end scenarios, providing ample, cost-effective computing resources, an easy-to-use, efficient, full-process AI platform;
    Globally, continue to work extensively with academia, industry and industry partners;
    Introduce AI thinking and technology into existing products and services to achieve greater value and competitiveness;
    Apply AI to optimize internal management, aim at massive operation scenarios, and greatly improve internal operation efficiency and quality;
    The launch of these new products by Huawei has caused extensive discussion in the industry. Coincidentally, the author also saw the exposure of Huawei's Arm server chip related products from insiders.

    Huawei 7 nm

    Arm server chip exposure

    Recently, Huawei officially disclosed its new generation of Arm server chip Hi 1620.

    According to informed sources, the semiconductor industry observers, Huawei's Arm server chip is designed independently based on the Arm V8 architecture, using the most advanced 7nm process in the industry. It is understood that Huawei provides 32, 48 and 64 core versions on this chip, up to 2.6/3.0Ghz, and can support PCIE 4.0 & CCIX.

    Huawei said that this is the industry's first 7nm Arm server chip supporting PCIE4.0. From Huawei's PPT, we can see that the Hi 1620's 48-core version of the CPU and Intel Skylake 8180's SPECint performance is equivalent, but in terms of power consumption will be 20% lower than the latter.

    Huawei Hi 1620 details

    As a wide-ranging enterprise, Huawei's Arm server chip has been developed for many generations.

    As you can see from wikichip, in 2015, Huawei introduced its first-generation Arm server chip Hi 1610. The 16-core chip designed with Arm Cortex-A57 can only achieve 2.1Ghz.

    In 2016, at the China Twelfth Five-Year Innovation Achievement Exhibition, Huawei exhibited its first ARM platform server “Taishan”, equipped with a self-developed ARM architecture 64-bit processor “Hi1612”, built using TSMC's 16nm process. , compatible with ARMv8-A instruction set. Huawei said that in addition to the storage unit, the processor has complete independent intellectual property rights and can be applied to big data analysis, shared cloud, information search and other fields, and has been tested in Alibaba.

    In 2017, Huawei introduced the HI 1616. The 32-core chip designed with Cortex-A72 has a maximum frequency of 3Ghz and then this year's Hi 1620. It can be seen that although Huawei has not publicized its Arm server chip, it has maintained an annual update frequency in the past few years.

    Huawei Arm server chip series

    Considering the influence of Huawei itself in mobile phones, cloud and storage, the arrival of this Arm server product is a further improvement of its own industrial chain for Huawei itself. Can provide customers with customized, comprehensive and controllable one-stop service.

    Zooming into the entire Chinese integrated circuit industry, Huawei's product line may be able to take a new path in the server chip market that Intel controls. But there is no doubt that this will face challenges from multiple competitors at home and abroad.


    Arm server chip market

    In recent years, with the increasing market share of Intel server chips, the rise of domestic independent controllable demand, Marvell acquired the establishment of Cavium, Huaxintong, Qualcomm's fading out, and the Arm server chip market has been surging. Although some people are withdrawing from the beginning, under the impetus of Arm, there are also new players entering this market. Huawei is one of them. As mentioned above, from the perspective of Huawei's business, the Arm server chip business is a supplement to the industry chain for them.

    In addition to Huawei, domestic Feiteng, Huaxintong, and American Ampere are also important players in the Arm server market.

    First look at the Feiteng aspect.

    Earlier, Dou Qiang, the chief scientist of Tianjin Feiteng Information Technology Co., Ltd., mentioned in an interview with the media industry observation that Feiteng launched the Feiteng FT2000+ processor in 2017. The chip built using the 16nm process has 64 cores and main The frequency can be 1.8-2.3GHz, and the measured performance of the standard spec test is comparable to that of the Intel Xeon processor introduced in 2013. Feiteng also completed the work related to server storage, database and middleware adaptation.

    In Dou Qiang's view, the performance of this processor is quite different from that of Intel's products. Even their products are single-channel design, which cannot meet the large-scale design requirements. But Feiteng will expand it two or even eight in the future to match the processor needs of high-end servers.

    Gu Hong, general manager of Feiteng, said before that Feiteng's CPU is based on ARM technology architecture, but the code part including CPU calculation module is independently developed by the company for many years. This allows Feiteng to have higher autonomy in the autonomous control of this series of products.

    I came to Huaxintong, a company jointly established by the Guizhou government and Qualcomm, focusing on the Arm server chip.

    According to a report by Phoenix Technology in May this year, Huaxintong's first server chip, “Huaxin No.1”, has been successfully produced at the end of 2017 and will be launched in the second half of this year. The second generation product they developed, "Huaxin No. 3", is currently under development.

    According to reports, this server chip has only half a bank card, integrating about 1 billion transistors and more than 2,800 pins, and the chip process is 10 nanometers. The built-in independent security module greatly enhances the chip safety factor, which is a highlight of "Huaxin No. 1". It can be applied to high-performance computers to play the role of processing large amounts of data quickly and in a timely manner.

    As for Ampere, it was founded by former Intel executive Renee James. In an interview with Ms. James before the semiconductor industry observation, she mentioned that Ampere's core team mostly comes from chip giants such as Intel and AMD. Most of the company has very rich experience in server hardware and software, they are on the server. The understanding of chips and software is quite deep, which makes them an emerging force in the field of Arm servers.

    In September of this year, Ampere introduced the 16nm process processor built by the company's first 64-bit Armv8-A architecture for the data center. Their 32-core Armv8-A processor is designed in Turbo mode. The frequency is up to 3.3 GHz. The processor has been chosen by Lenovo and several other original design manufacturers (ODMs).

    According to them, this processor has excellent total cost of ownership (TCO) value, powerful computing performance and memory capacity, and rich I/O to handle cloud workloads, including big data, web tiers, and in-memory databases. .

    Ampere also announced future multi-generation product roadmaps, including next-generation 7nm products. The product will offer single-socket and multi-socket options and will be available in 2019, which will be used for future ultra-large-scale cloud computing and edge computing.

    As can be seen from the above, Huawei's leading edge in the Arm server chip is ahead of its global competitors.

    to sum up

    Although Huawei's Arm server chip has so far dominated, we can see that Intel's server ecosystem, which has been built for decades, cannot be shaken. However, Huawei relies on its chip design experience accumulated over the years, and has been in the field in the past year. Coupled with Huawei's own accumulation of AI chips, ISP chips, mobile phone SoCs and other various chips, terminals and applications, Huawei will play an important role in the Arm server market in the future.

    As for the future, it depends on how Arm combines the major chip suppliers and software vendors to work together in this field.

    半导体行业观察 2018-10-11 10:39:12
    来源:内容由 微信公众号 半导体行业观察 (ID:icbank) 李寿鹏 原创,谢谢。



    据华为轮值总裁徐直军介绍,华为这次推出的两款芯片分别是昇腾910和昇腾310,这都是华为基于其自研的达芬奇架构打造的两款新品。其中昇腾910是当前单芯片计算密度最大的AI芯片。该产品采用7nm工艺打造,最大功耗做到350w,其他参数也是表现优越:在半精度 (FP16)下,可以做到256 TeraFLOPS,在整数精度 (INT8)下,更是做到了512 TeraOPS,另外,该款芯片还支持128 通道的全高清视频解码器(H.264/265)。


    徐直军表示,华为昇腾910在与英伟达和谷歌的芯片对比时,也拥有强大的领先优势。芯片将在20 19年Q2到来,这会在云端给华为带来训练和逻辑服务系列的强大支持,冲破现在市场被TPU和英伟达垄断的局面。


    另外,华为还发布了昇腾310,按照徐直军的说法,华为这款极致高效计算低功耗的AI SoC是针对边缘AI而推出的产品。作为一款同样采用达芬奇架构的芯片,华为昇腾310采用了12nm FFC工艺制造,在半精度 (FP16)情况下,可以做到8 TeraFLOPS,在整数精度 (INT8) 下,则能做到16 TeraOPS,还能支持16 通道全高清视频解码器 - H.264/265,而其最大功耗只有8W,这款芯片现在就已经能够给客户提供全方位的支持。


    华为方面表示,他们基于统一、可扩展架构的系列化 AI IP和芯片昇腾拥有nano、tiny、mini、lite和max五个系列,能提供横跨全场景的最优TOPS/W支持。




    • 在计算视觉、 自然语言处理、 决策推理等领域构筑数据高效(更少的数据需求)、 能耗高效(更低的算力和能耗),安全可信、自动自治的机器学习基础能力;
    • 打造面向云、 边缘和端等全场景的、 独立的以及协同的、 全栈解决方案, 提供充裕的、 经济的算力资源, 简单易用、 高效率、 全流程的AI平台;
    • 面向全球, 持续与学术界、产业界和行业伙伴广泛合作;
    • 把AI思维和技术引入现有产品和服务, 实现更大价值、更强竞争力;
    • 应用AI优化内部管理, 对准海量作业场景, 大幅度提升内部运营效率和质量;



    日前,华为正式对外披露了其新一代的Arm服务器芯片Hi 1620。

    据知情人士告诉半导体行业观察记者,华为这颗Arm服务器芯片是基于Arm V8 架构自主设计的,使用当前业界最先进的7nm工艺打造。据了解,华为在此芯片上提供32、48和64核的版本,最高支持2.6/3.0Ghz的主频,能够支持PCIE 4.0&CCIX。

    华为方面表示,这是业界第一颗支持PCIE4.0的7纳米Arm服务器芯片。从华为的PPT中我们可以看到,Hi 1620的48核版本的CPU和英特尔Skylake 8180 的SPECint 性能相当,但在功耗方面会比后者低20%。

    华为Hi 1620的细节


    从wikichip可以看到,2015年,华为推出了其第一代Arm服务器芯片Hi 1610,这个采用Arm Cortex-A57设计的16核芯片主频最高只能做到2.1Ghz。


    2017年,华为又推出了HI 1616,这个采用Cortex-A72设计的32核芯片最高主频可以做到3Ghz,再到今年Hi 1620。可以看到,虽然华为并没有大肆宣传其Arm服务器芯片,但是在过去的几年也都保持每年一款的更新频率。















    至于Ampere,则是由Intel前高管Renee James创立的。在半导体行业观察之前对James女士发起的专访中她提到,Ampere的核心团队大部分来自Intel和AMD这些芯片巨头,公司的大多数人在服务器的软硬件领域拥有非常丰富的经验,他们对服务器芯片和软件的理解相当深入,这就使得他们成为Arm服务器领域的新兴势力。

    在今年九月,Ampere推出了该公司旗下面向数据中心的第一代 64 位 Armv8-A架构的,16nm工艺打造的处理器,这款他们设计的 32 核 Armv8-A 处理器在Turbo 模式下主频高达 3.3 GHz。处理器已获得联想及其他几家原始设计制造商 (ODM) 的选择。

    按照他们的说法,这款处理器具有优秀的总体拥有成本 (TCO) 价值、强大的计算性能和内存容量以及丰富的 I/O,用来处理云工作负载,包括大数据、Web 层以及内存数据库。

    Ampere 还公布了未来多代产品路线图,包括下一代 的7nm 产品等。这款产品将提供单插口和多插口选项,并于 2019 年上市,这将用于将来的超大规模云计算和边缘计算。





    Huawei Shengteng 910 Introduction

    Xu Zhijun said that Huawei's Shengteng 910 also has a strong lead when compared with NVIDIA and Google's chips. The chip will arrive in Q2 in 2019, which will give Huawei a strong support for the training and logic service series in the cloud, breaking through the current market monopoly by TPU and Nvidia.
    • Thanks Thanks x 9
  2. somsak

    somsak FULL MEMBER

    Jun 27, 2014
    +1 / 1,449 / -1
    A Link to Source, please
  3. grandmaster

    grandmaster FULL MEMBER

    Dec 20, 2011
    +0 / 300 / -0
    Shhhh, shut up. Don't say anything about this. If not, Uncle Trump and Viet bois cannot hold their jealousy anymore. They will start venting their jealousy on Huawei by making fake news about Huawei and lobbying against Huawei more and more.
    • Thanks Thanks x 2
  4. oprih


    Dec 21, 2015
    +0 / 5,612 / -21
    Nice, btw the article is too complicated, the americans and their slaves are too dumb to understand it.
    • Thanks Thanks x 2