The first domestic 7nm general-purpose GPU is here! 24 billion transistors, commercial soon

Chipsets reported on March 31 that, just now, the first domestically controllable 7nm cloud general-purpose GPU was officially released in Shanghai.

This chip is the flagship 7nm GPGPU cloud training chip BI developed by Shanghai start-up company Tianshu Zhixin. Its accelerator card is also released in physical form and will soon enter mass production and commercial delivery.

Tianshu Zhixin was established in 2015 and officially launched the BI R&D design in 2018. The chip was successfully “lit” in May 2020, film production in November, and December.

It is reported that BI chips can provide nearly twice the peak performance of mainstream manufacturers’ products with 50% of the chip area of ​​competing products and lower power consumption.

The first domestic 7nm general-purpose GPU is here! 24 billion transistors, commercial soon

Ni Guangnan, Academician of the Chinese Academy of Engineering, Vice Chairman of the Foreign Affairs Committee of the 12th National People’s Congress, Chairman of the “One Belt and One Road” International Think Tank Expert Committee of the Chinese Academy of Social Sciences, and Chairman of the Landi International Think Tank Expert Committee, Dr. Zhao Baige, also shared in particular about the independence of science and technology at the meeting. Self-improvement insight.



24 billion transistors, peak computing power up to 147TFLOPS

Zheng Jinshan, senior vice president, chief scientist, and co-founder of Tianshu Zhixin, introduced the cloud training chip BI and product cards on the spot. The BI chip has high performance, versatility and flexibility.

As Tianshu Zhixin’s first flagship product, BI adopts TSMC’s 7nm FinFET process, 2.5D CoWoS package, based on a fully self-developed GPGPU architecture, accommodating 24 billion transistors, compatible with mainstream GPGPU ecology, and supporting mainstream deep learning frameworks.

 The first domestic 7nm general-purpose GPU is here! 24 billion transistors, commercial soon

▲Tianshu Zhixin BI product parameters

Specifically, the chip has built-in multi-precision data mixed training such as FP32, FP16/BF16, INT32/16/8, and its single-core peak computing power reaches 147TFLOPS under FP16 accuracy.

At present, the measured data of BI products basically conform to the design plan.

Tianshu Zhixin’s fully self-developed GPGPU architecture is a scalable hierarchical computing engine based on the SIMT architecture. The self-defined rich instruction set supports scalar, vector and tensor operations, and supports the GPU general parallel programming model, which can effectively interface with the existing software ecosystem. It is easy to expand to support new algorithms and application areas, and it is easy for users to migrate easily.

  The first domestic 7nm general-purpose GPU is here! 24 billion transistors, commercial soon

▲Tianshu Zhixin BI Architecture

BI chip also provides 1.2TB/s HBM2 memory bandwidth, 32GB storage capacity, inter-chip interconnect bandwidth up to 64GB/[email protected] PCIe4, and supports virtualization.

Its standard acceleration module group (OAM) based on the Open Computing Project Group (OCP) supports a system solution with a maximum single card of 300W-450W, and cooperates with the OAM server to further improve the overall performance of data processing.

 The first domestic 7nm general-purpose GPU is here! 24 billion transistors, commercial soon

▲Tianshu Zhixin BI OAM Product Card

In order to facilitate the use of developers, Tianshu Zhixin has also created a software stack compatible with mainstream deep learning frameworks, which can help users achieve painless migration.

At the same time, Tianshu Zhixin software stack combines hardware performance to provide fine-grained optimization and greater computing power support for applications such as HPC and blockchain, including resource management and monitoring plug-ins that can interface with various Internet application platforms, and deep learning acceleration libraries , Compatible with multi-language compilers, tuning tools, etc.

 The first domestic 7nm general-purpose GPU is here! 24 billion transistors, commercial soon

▲Tianshu Zhixin Software Stack

After mass production, Tianshu Zhixin BI chip and product card can provide computing power support for high-load work such as AI training and reasoning, cognitive AI, high-performance data analysis, genome research, financial predictive analysis, and serve education and the Internet. , Finance, autonomous driving, medical, security and other related industries.


  Ni Guangnan: The “Chinese system” in the cyberspace industry continues to grow and develop, but the pain points still exist

During his speech, Cai Quangen, chairman of Tianshu Zhixin, said: “The long-term goal of Tianshu Zhixin is to steadily climb on the domestic high-end independent high-end computing power chips and eventually compete with international leading manufacturers.”

“This goal is not easy to achieve, but we will not seek imitation shortcuts and speculative corner overtaking.” Cai Quangen said.

  The first domestic 7nm general-purpose GPU is here! 24 billion transistors, commercial soon

▲Cai Quangen, Chairman of Tianshu Zhixin, delivered a speech

Ni Guangnan, an academician of the Chinese Academy of Engineering, also delivered a speech.

He mentioned that since the 18th National Congress of the Communist Party of China, my country’s information industry has achieved leapfrog development. At present, some independent and controllable core technologies in the field of cyberspace have become available, and are developing in the direction of ease of use and ease of use. The “Chinese system” in the field continues to grow and develop.

However, shortcomings and pain points still exist. For example, the most complex circuit chips in the two information systems, CPU and GPU, have increasingly become shortcomings that must be broken through in the entire industry chain.

The good thing is that we have gradually accumulated more and more talents and technologies in the shortcomings. We can use our own system advantages, concentrate our efforts on major tasks, make full use of the “new infrastructure” opportunities, and quickly learn from each other.

  The first domestic 7nm general-purpose GPU is here! 24 billion transistors, commercial soon

▲Academician Ni Guangnan of the Chinese Academy of Engineering delivered a speech

“Tianshu Zhixin’s BI chip has always adhered to both independent control and open innovation. It currently has its own technical system and ecology, and its performance is also very good. It has received strong support from the government, many investors and partners.”

Academician Ni Guangnan said that he expects Tianshu Zhixin to gradually establish its own standards in the industry, gradually develop and grow, and be able to compete with international enterprises on the same stage, and contribute to building a technologically independent and self-reliant country.


Zhao Baige: Must maintain a balance between independent research and development and openness

Dr. Zhao Baige is the vice chairman of the Foreign Affairs Committee of the 12th National People’s Congress, the chairman of the Chinese Academy of Social Sciences’ “Belt and Road” international think tank expert committee, and the chairman of the Landi international think tank expert committee.

She mentioned that she started to pay attention to Tianshu Zhixin very early, which is also the enterprise that Landi International think tank platform focuses on fostering and promoting.

In her view, the successful development and launch of physical products by Tianshu Zhixin BI has opened a breakthrough for the development and application of my country’s independent high-end mainstream general-purpose chips, and also provided a breakthrough for the development of domestic computing power-intensive industries and the implementation of artificial intelligence. More and better choices, but also a layer of security.

  The first domestic 7nm general-purpose GPU is here! 24 billion transistors, commercial soon

▲Dr. Zhao Baige delivered a speech

Dr. Zhao Baige said that after very in-depth research, they found that they must know oneself and know each other, and must compete and cooperate. The United States has a very good foundation in scientific and technological innovation, but China has no other regions to compare in terms of market application. We must Looking for a path of self-development, and at the same time recognize fresh development opportunities.

What to do if you want to continue to develop? Dr. Zhao Baige believes that some problems should be solved by the government, and some problems should be solved by enterprises.

The first is the impact of the international situation. We must recognize who we are, where we are going, who they are and where they are going. We must maintain a balance between independent research and development and openness, so that we can truly achieve breakthroughs in the world and promote China’s knowledge economy industry.

At the same time, Dr. Zhao Baige mentioned that China still has a big gap in the underlying technology (and the United States). The United States has attracted talents from all over the world and has a certain ability to correct errors. After problems are discovered, they will quickly adjust and develop the underlying technology and strategy. , Talents are closely linked.

In addition, the communication and cooperation between government and enterprises are very important, and management innovation is worth studying. Political achievement is not only the GDP content, but also the GDP quality that is closely linked to the knowledge economy, including the support and protection of talents. We must be more open-minded. Recruiting talents from all over the world.

At the same time, we must also pay attention to the protection of intellectual property rights and integrate corporate development with the national grand strategy. Dr. Zhao Baige hopes that the release of Tianshu Zhixin will also trigger new ideas and considerations on resource combination.


▲Tianshu Zhixin teamed up with Inspur to create OAM AI server


In the past year, GPGPU startups have become “golden beasts”

At the beginning of this month, Tianshu Zhixin announced the completion of 1.2 billion yuan of C round of financing, and plans to use the funds to further accelerate the marketization, commercialization and scale of cloud training and inference chips for 5G technology needs.

In addition to Tianshu Zhixin, many domestic GPGPU startups focusing on the cloud chip track have announced high financing in the past year.

Why have local GPGPU startups become popular in the capital market?

This is dictated by the times. As artificial intelligence technology moves from universities to industry, the demand for GPGPU, which is good at general parallel computing, has soared. The international GPU giant NVIDIA has mastered most of the cloud AI training chip market, and there is almost no opponent in the world that can compete with it.

On the one hand, this has led to the lack of bargaining power in the downstream industry, and on the other hand, it has also led to security concerns. In the context of endless international trade disputes, the independent and controllable key links of the chip semiconductor industry are imminent, and many countries are planning core technologies and localization of the supply chain. .

The GPGPU, whose domestic market capacity and demand continue to rise, is undoubtedly a technological highland that must strive for independence and control.

Under the tide of new infrastructure, the development of industries such as artificial intelligence, cloud data centers, intercity transportation, new energy vehicles, and the popularization of 5G applications are constantly catalyzing the demand for domestic cloud AI computing power.

However, the main reason why the GPGPU market has been controlled by NVIDIA for a long time is that its technical threshold is high and the ecological layout is perfect. Talents who have experienced the complete process of advanced process research and development and mass production are very scarce.

This is a winner-takes-all field. NVIDIA, which has abundant talents and market resources, continues to accumulate and attract developers through CUDA, thus forming an increasingly indestructible ecological barrier.

Therefore, most startups that choose the GPGPU track are very cautious when taking the first step. Although they target NVIDIA, they mainly focus on the performance in a specific direction of Yoha to differentiate their advantages, and then gradually Expand product boundaries.

At the same time, considering that the vast majority of developers are accustomed to using NVIDIA CUDA, these companies are focusing on developing convenient and easy-to-use their own software tool chains on the one hand, and on the other hand, they are basically compatible with the mainstream GPGPU ecosystem.


Conclusion: The domestic GPGPU commercial future is still full of challenges

Nowadays, domestic GPGPU players are all around the world, and many startups, including Tianshu Zhixin, continue to make efforts in financing and product commercialization. As more players’ chips move from research and development to landing, domestic GPGPU startups will gradually transition from the competition of capital and talent to a new contest of product competition and market competition.

The road to local self-developed chips is often difficult and full of risks. For these start-up teams, in addition to developing chip products that can satisfy customers, how to find a living space in a market dominated by giants and how to build their own ecology… are all challenges that need to be overcome on the road to future commercialization. .


The Links:   NL4827HC19-05B FF450R12ME4

Leave a Reply

Your email address will not be published. Required fields are marked *