Technology giants are hunting for "AI chips", and mainstream manufacturers at home and abroad are combing

The launch of Gemini is considered by Google to be the most critical technological innovation of the decade, and it has been pre-trained on different modes using Google-designed TPUs v4 and v5e chips from the beginning of the design.

The AI chip is the core, the large model is the platform, and the application is the product.

Tech giants know that only by getting the right to speak on AI chips can they stand out in this artificial intelligence war.

Foreign and domestic manufacturers have already entered the market.

Google - TPU

Compared to CPUs and GPUs of the same period, TPU can provide a 15-30x performance increase and a 30-80x efficiency (performance/watts) boost.

This time Google released the TPU v5p, which is Google's most powerful and efficient TPU chip to date. Designed to improve performance, scale, and flexibility, TPU v5p is 2.8x faster than TPU v4 and 1.9x faster at training embedded-dense models. At the same time, TPUv5p improves FLOPS by two times and memory bandwidth by three times compared to TPU v4.

Zhongji InnoLight: Google's core supply chain, VFR8 optical modules with circulators have strong profitability and large share advantages. Google's core provider of data center optical transceivers, benefiting from Google's investment in cloud infrastructure.

Tengjing Technology: We are currently sending samples of the circulator to Cloudlight, the supplier of Google's optical modules

Borche Technology: Acquired Changxingsheng to become a core supplier of Google's data centers, which is expected to benefit from Google's optical switch needs.

T&S: Provide connector products to Google. The customers are mainly from Google, and they have begun to deploy in Google, benefiting from Google's data center optical module needs.

AMD--MI300

With over 150 billion transistors and 2.4 times the memory of Nvidia's H100 chip, the MI300X is claimed to be on par with the H100 chip at training large language models, but better at inference.

The Instinct MI300X accelerator has been shipping since the launch date, and the Instinct MI300A accelerator has entered mass production.

Tongfu Microelectronics: AMD's largest packaging and testing supplier, the company has formed a strong alliance model of "joint venture + cooperation" with AMD through mergers and acquisitions, and has packaging and testing projects involving AMDInstinctMI300 and Rx7900mGPU.

VeriSilicon: In April 2022, the company officially joined the UCIe Industry Alliance (jointly established by AMD and other ten companies in March 2022 and proposed an open chiplet interconnection specification), and the company continues to promote the industrialization of chiplet technology and is expected to enter the core chiplet industry chain.

Ruijie Network: Cooperate with AMD to develop GPU solutions.

CLP Port: Distribution of AMD EPYC series CPUs.

Aoshikang: AMD's PCB supplier.

Victory Giant Technology: AMD's PCB supplier.

Nvidia - A100/H100/H200

A100 and H100 (mainstream in the market), as well as the recently released new generation of H200 chips.

Compared with the previous generation A100 and H100, the main change of H200 is memory, becoming the first GPU to use HBM3e (high-frequency wide memory), which increases the bandwidth from 3.35TB/s to 4.8TB per second of H100, an increase of 1.4 times, and the total memory capacity of H100 is increased from 80GB to 141GB, the capacity is increased by 1.8 times, and the inference energy consumption is greatly reduced.

Zhongji InnoLight: Nvidia supplier, providing 800G optical modules.

Tianfu Communication: The second phase of the optical engine has started, and Nvidia has a high revenue ratio in 23 years.

Aoshikang: Supplying PCBs to NVIDIA through a supply system.

Shanghai Electric Co., Ltd.: There is a hash board material number that is being certified by NVIDIA.

Robotec: Ficontec, a joint stock company, has been supplying equipment to NVIDIA.

Boke New Materials: Exclusive supplier of soft magnetic powder cores for NVIDIA chip inductors.

微软--Maia 100 /Cobalt 100 Arm

On November 15, Microsoft announced two chips at the Ignite conference in Seattle, one for AI and one for cloud computing. The AI chip, called the Maia 100, was released by Microsoft to compete with NVIDIA's highly sought-after AI graphics processing unit. The second is the Cobalt 100 Arm chip, which is aimed at general-purpose computing tasks and will compete with Intel processors for the market.

Amazon – Training chip Trainium/Inferentia chip

Of the three — Amazon, Microsoft, and Google — Amazon is the only cloud provider that offers both training and inference chips in its servers. Amazon Web Services currently has three self-developed chip production lines, namely the general-purpose chip Graviton, the dedicated AI chip Trainium (training) and Inferentia (inference), and Nitro.

Huawei--Ascend 910/Ascend 310

At present, it mainly includes two processors, the Ascend 910 (for training) and the Ascend 310 (for inference), which use its own Da Vinci architecture.

Runhe Software: The company is Huawei's in-depth war loss partner, and has developed a variety of industry terminal solutions on the Kirin series chip platform.

Digital China: The company is an ecological partner of Kunpeng and Ascend, and the company is an in-depth strategic partner of Huawei, and the company has released the Shenzhou Kuntai AI inference server, with "Kunpeng + Ascend" as the core, which can provide computing power of 128 processing cores, and can support up to 8 Huawei Atlas 300 inference cards, provide 256GB of inference cache, and provide AI computing power of up to 704 TOPS INT8.

Fiberhome Communications: 87% of the shares of Yangtze River Computing. As an important hardware partner of Ascend, YMC has launched a variety of AI servers, such as G210K V2 edge, G220K V2 inference, and G440K V2 training, based on Ascend AI, to meet the business needs of AI infrastructure construction.

Based on the Kunpeng processor and Ascend processor, Xiangjiang Kunpeng, a holding subsidiary of the company, has successively released dozens of Zhaohan AI products, such as AI inference, AI training, AI small stations, AI clusters, and intelligent edges, and has joined hands with Huawei to participate in the construction of AI computing centers across the country.

Tongfang Co., Ltd.: On February 16, 2023, Tongfang Computer Co., Ltd., a subsidiary of the company, jointly launched an AI server based on Ascend technology with Huawei, and signed a cooperation agreement with Beijing Ascend Artificial Intelligence Computing Center.

Haiguang Information - Haiguang series, deep computing series AI chips

Haiguang Information: Background of the Chinese Academy of Sciences, the company's products include Haiguang General Purpose Processor (CPU) and Haiguang Coprocessor (DCU). Haiguang CPU series products are compatible with x86 instruction set and international mainstream operating systems and application software, with rich software and hardware ecology, excellent performance, safety and reliability, and have been widely used in important industries or fields such as telecommunications, finance, Internet, education, and transportation.

Cambrian - Siyuan series Ai chips

Cambrian: Siyuan 370 is Cambrian's third-generation cloud product, using 7nm process technology, and is Cambrian's first AI chip with Chiplet technology, with a maximum computing power of up to 256TOPS (INT8). Cambrian's new generation of cloud intelligent training Siyuan 590 chip has not yet been released, but it has received a lot of attention and discussion, especially the high hope that it will replace A100 to a certain extent in large model training and inference tasks.

Loongson Zhongke - Loongson series chips

Loongson Zhongke: Loongson 3A6000 adopts the instruction system and architecture independently designed by the mainland, and is a new generation of general-purpose processor independently developed and controllable by the mainland, which can run a variety of cross-platform applications and meet all kinds of large-scale and complex desktop application scenarios.

Suiyuan--Si series chips

Tencent background, the full range of computing power product lines of the Diosi series of chips, Yunsui training and inference acceleration cards, and Yunsui intelligent computers. At present, Suiyuan has deployed a kilocalorie-scale AI training computing power cluster for large scientific research institutions, and has successfully landed, and Suiyuan has also cooperated with Tencent to play a role in OCR text recognition, intelligent interaction, intelligent conference, etc., with more than twice the performance of similar products in the industry, and has a high advantage in cost performance.

Hongxin Electronics: After the company signed a strategic cooperation agreement with Suiyuan Technology, it carried out multi-level and in-depth collaborative cooperation.

China's local AI chip companies also include Baidu Kunlun Chip, Biqian Technology, Tiantian Zhixin, Muxi, etc.

I will be in the Official Account: Haihan Finance Daily updates the latest Huawei concepts, innovative weight loss drugs, digital economy, ChatGPT, AI computing power, CPO/silicon photonic chips, big data, 6G satellites, data elements, new medical infrastructure, integrated die-casting, automotive intelligence, lidar, HUD, automotive chips, air suspension, L3 intelligent driving, PET copper foil, nano battery, 800V high voltage, photovoltaic HJT, TOPCON, perovskite, photovoltaic XBC, BIPV, IGBT chips, Silicon carbide SIC, CTP/CTC/CTB battery, 4680 battery, industrial machine tool, sea breeze flexible straight high voltage, new energy vehicle high-voltage fast charging, high nickel ternary, carbon fiber, PET aluminum foil, PET copper foil, air source heat pump, new materials, traditional Chinese medicine innovative drugs, traditional Chinese medicine formula granules, rural revitalization, lithium mines, vanadium flow batteries, sodium-ion batteries, distributed energy storage, centralized energy storage, The latest hot topics such as antigen testing are excavated, and the future belongs to the structural market with high expectation difference, and it is fundamental to grasp the core track and the internal value logic expectation difference of individual stocks.

— END —

Like first and then watch, develop a Xi

Disclaimer: The pictures and data come from the Internet, and the reprint is only used for communication Xi, if there is any copyright problem, please contact the author to delete