Recently, the market research company Compass Intelligence released the latest research report. In the ranking list of the world's top 15 AI chip companies, the top three are Nvidia, Intel and IBM, and Huawei ranks 12th, becoming the TOP15. China's "only seedlings".
It is understood that the A-list in the AI Chipset Index of this report includes companies that provide software and hardware components for AI chipsets.
The AI chipset products include central processing unit (CPU), graphics processing unit (GPU), neural network processor (NNP), application specific integrated circuit (ASIC), field programmable gate array (FPGA), reduced instruction set computer (RISC) ) processors, accelerators, etc. Some chipsets target edge processing or devices, some target servers used in cloud computing, and others target machine vision and autonomous vehicle platforms. Some of these products are computing frameworks for AI, while others are AI training platforms.
The report also mentioned that in the past three years, in addition to its own research and development investment, a total of up to 60 billion US dollars has been invested in the field of artificial intelligence. We have seen that more than 1,700 startups are currently interested in AI chips. Of course, The industry's demand for AI chips is also increasing.
At the same time, in addition to the TOP15 ranking, Compass Intelligence also evaluated more than 100 chip companies, and 24 companies were shortlisted in the final ranking, including NVIDIA, Intel, IBM, Google, Apple, AMD, ARM, Qualcomm, Samsung, NXP, etc. It is worth noting that the Chinese company Huawei is still ranked 12th, Cambrian and Horizon are 22nd and 24th respectively.
In the Top24 list, a total of seven Chinese companies were shortlisted.
Huawei (HiSilicon) ranks 12th in this list;
MediaTek ranked 14th;
Imagination ranked 15th;
Rockchip ranked 20th;
Verisilcon ranked 21st;
Cambricon ranked 23rd;
Horizon ranked 24th;
Huawei's "Road to Core Manufacturing"
In October 2004, Huawei founded HiSilicon, its predecessor was Huawei Integrated Circuit Design Center, which officially opened Huawei's mobile phone chip research and development road.
2009年华为推出了第一款面向公开市场的K3处理器，定位跟展讯、联发科一起竞争山寨市场，华为自己的手机没有使用。因为K3产品不够成熟以及不适的销售策略，这款芯片并没有成功。，这也是国内第一款智能手机处理器。
2012年华为海思推出K3V2处理器，这一次用在了自家手机中，而且是定位旗舰的Mate 1、P6等机型。2012年手机处理器已经开启多核进程， K3V2成为了世界上第二颗四核处理器。
而后，麒麟910是海思的第一款SoC，如果说CPU是手机大脑，那SoC就是集成身体各种机能并给它们分配任务的系统，一个移动SoC除了CPU还包括基带（Baseband）、图形处理器（GPU）、数字信号处理器（DSP）、图像信号处理器（ISP）等重要模块。
从K3V2以来，部分华为手机特别是旗舰机一直使用自己的海思芯片，更重要的是，华为旗舰的绑定倒逼海思，必须迅速进步并且稳定供货。
我们可以看到，2014年华为的研发投入比A股400家企业的总和还多。2017年华为研发费用高达897亿人民币，大大超过苹果和高通。过去十年，华为投入的研发费用超过3940亿元，居于世界科技公司前列。
这样的成绩也就不足为奇了。
经过十几年的发展，2017年9月，华为在德国柏林国际电子消费品展览会（IFA）上正式推出其最新 AI 芯片 “麒麟970”（Kirin 970）。麒麟970采用行业高标准的 TSMC 10nm 工艺，在指甲大小的芯片上，集成了55亿个晶体管，功耗降低了20%，并实现了 1.2Gbps 峰值下载速率。麒麟 970集成 NPU 专用硬件处理单元（寒武纪IP），创新设计了 HiAI 移动计算架构，其 AI 性能密度大幅优于 CPU 和 GPU。相较于四个 Cortex-A73核心，处理相同 AI 任务，新的异构计算架构拥有约50倍能效和 25 倍性能优势。
而且，华为第二代AI芯片海思麒麟 980也将在本季度正式量产，采用台积电 7nm 制程工艺。这款处理器将配置第二代 NPU，在前代的基础上，支持更多的场景应用，NPU 的性能提升 2 倍以上。
中国“造芯运动”
5月3日，寒武纪在上海发布了新一代终端 IP 产品，采用7nm工艺的终端芯片Cambricon 1M、首款云端智能芯片MLU100以及搭载了MLU100的云端智能处理计算卡。
第三代机器学习终端处理器1M其性能比此前发布的寒武纪1A高10倍。配置方面，寒武纪1M使用台积电7nm工艺生产，其8位运算效能比达5 Tops/watt（每瓦 5万亿次运算）。寒武纪提供了2Tops、4Tops、8Tops三种尺寸的处理器内核，以满足不同场景下不同量级智能处理的需求。
而MLU100采用寒武纪最新的MLUv01架构和台积电16nm工艺，可工作在平衡模式（主频 1Ghz）和高性能模式（主频1.3GHz）两种不同模式下，等效理论峰值速度则分别可以达到128万亿次定点运算和166.4万亿次定点运算，而其功耗为80w和110w。
寒武纪介绍，MLU100云端芯片同样具备高通用性，可支持各类深度学习和常用机器学习算法，他们还提出“端云协作”的理念，也就是说，MLU100云端芯片可以和寒武纪1A/1H/1M系列终端处理器进行适配，协同完成复杂的智能处理任务。
而在此前，阿里巴巴、地平线、云知声、Rokid等中国高科技公司都宣布加入“造芯运动”，就阿里巴巴而言，他们正研发神经网络芯片Ali-NPU，这款芯片性能将是目前市面上主流 CPU、GPU 架构 AI 芯片的 10 倍，而制造成本和功耗仅为一半，性价比超过 40 倍，一天之后，阿里巴巴再度宣布全资收购中天微，而后者是中国大陆唯一的自主嵌入式?CPU IP Core 公司。
As early as the end of last year, Horizon released embedded artificial intelligence chips - Journey 1.0 processor for intelligent driving and Sunrise 1.0 processor for intelligent cameras. Yu Kai believes: "The future that Horizon sees is artificial intelligence processors, which are actually the commanding heights of our country's technological competitiveness. If China's artificial intelligence industry wants to take off in the future, it must have aero-engines. What? It must be an artificial intelligence processor."
There are still many startup companies joining this battle. Both Yunzhisheng and Rokid have announced the completion of chip research and development. Yunzhisheng is about to release an AI chip, which is based on Unisound's AI instruction set and DSP instruction set, combined with voice application scenarios. , a new chip architecture that integrates microphone array signal processing, speech recognition and speech synthesis.
According to reports, this AI chip uses a multi-level-multi-group-multi-port Memory architecture to ensure the improvement of on-chip data bandwidth and reduce chip power consumption while ensuring computing efficiency through a programmable interconnect matrix between computing units. In terms of architectural flexibility, the main control CPU is connected to the internal RAM of the AI accelerator through the Scratch-Pad, which provides an efficient data channel between the CPU and the AI accelerator, so that the CPU can perform secondary processing on the AI accelerator operation results. In addition, the programmable interconnection matrix structure connecting each operation unit provides the function of expanding operation instructions, thereby further improving the flexibility and scalability of the hardware architecture. The remaining exploration of chip architecture, including multi-level and multi-mode wake-up, from energy detection to human voice detection to wake-up word detection, and customized Power Domain for voice devices and usage scenarios, minimize chip power consumption. (From: NetEase Intelligence Editor: Wang Chao_NT4133)
Article source: Netease Intelligence