Tianshu Zhixin Tiangai 100 product card took the lead in mass production and delivery, and made a new breakthrough in the field of high-end GPGPU in my country

Tianshu Zhixin Tiangai 100 product card took the lead in mass production and delivery, and made a new breakthrough in the field of high-end GPGPU in my country

On October 29, Tianshu Zhixin announced that the company’s fully self-developed and domestic first cloud 7nm GPGPU product card – “Tiangai 100” has officially entered mass production, and will provide the industry with leading and strong AI computing power in the future. Tianshu Zhixin products are the first to enter the commercialization process.

At present, the performance of Tiangai 100 is comparable to that of mainstream products in the industry. The product performance has reached and met the design goals of data centers, servers and other fields, and has achieved low-cost migration. In addition, the company has further optimized the product through early testing and user adaptation feedback, and has actually entered mass production. This is a key point for the domestic GPU track to move from the laboratory to the ground, and it is also another major breakthrough in the field of high-end GPGPU in my country!

In this regard, Mr. Diao Shijing, Chairman and CEO of Tianshu Zhixin, said: “Tianshu Zhixin has always adhered to the GPGPU strategy, continued to explore the GPGPU market and launched high-quality products that meet the needs of the market and users. We have the world’s top digital integrated circuit design and basic software. The team of design scientists has always insisted on taking one step at a time and doing things well in a down-to-earth manner. The official entry of Tiangai 100 product card into mass production is a new milestone for the entire industry. We will continue to combine the company’s first-mover advantage and continue to Focus on the pain points of customer needs, accelerate the process of product commercialization, and fully prepare for future product iterations.”

The difficulty of GPGPU design is obvious to all in the industry. From the software and hardware architecture of the product, the design of the computing core, to the development of the compiler, driver, and function library, every link requires extremely solid research and development skills. Easy. Usually, the front-end and back-end design of a high-end chip in the industry takes 1-3 years. After the design is completed, the tape-out process takes 3-6 months, and there is a risk of tape-out failure and all over again. Even if tape-out is successful, it still takes 3 to 12 months of product testing and tuning before mass production can finally start. In summary, a high-end self-developed chip generally takes 2 to 5 years to go from project approval to real mass production.

“For GPGPU chips with high computing power, in addition to the very professional and extremely demanding requirements for cutting-edge technology in the design and production of the chip itself, the underlying hardware programming is also very complicated.” Zhixin is the first provider of GPGPU cloud chips and super computing power systems in China, leading the most cutting-edge trend in the field of high-performance computing represented by AI. The colleagues of Tianshu Zhixin have always maintained the professional spirit of excellence and polishing products. In just three years, the product has successfully moved from design to mass production, which is all due to the cooperation of the company’s complete GPGPU team and the strong support of investors, customers, governments and partners.”

Tianshu Zhixin officially launched the GPGPU chip design in June 2018, which is compatible with mainstream development frameworks. During the test demo, the film was successfully returned after a tape-out, and the adaptation with the client server was started until it was officially released in March this year. Under the circumstance of limited personnel, the chip was completed in only about a year and a half. the design of.

The chip BI lived up to expectations, showing excellent performance. The BI chip adopts a 7-nanometer advanced process, accommodates 24 billion transistors, and adopts 2.5D CoWoS wafer packaging technology. It supports multi-precision data hybrid training such as FP32, FP16, BF16, and INT8, and the single-core computing power is [email protected] per second. With 1/2 the chip area and lower power consumption of similar products, it provides nearly twice the performance of mainstream manufacturers’ products. This is enough to reflect the research and development strength and pragmatic attitude of Tianshu Zhixin.

According to IDC, the GPU server market in China will reach US$6.4 billion by 2024. If the domestic GPU can achieve a 30% share in 2024, it can gain a market space of 2.2 billion US dollars. Therefore, the future demand of GPGPU in China is extensive and crucial to the development of the national economy.

As the first provider of GPGPU cloud chips and super computing power systems in China, Tianshu Zhixin will always be based on the domestic market, with the mission of “becoming an enabler of intelligent society”, and continue to explore and move forward. Efforts will be made to further promote the development of the domestic GPGPU industry and contribute an important force to the development of my country’s chip industry.

About Tianshu Zhixin

Shanghai Tianshu Zhixin semiconductor Co., Ltd. (referred to as “Tianshu Zhixin”) officially launched the design of 7-nanometer general-purpose parallel (GPGPU) cloud computing chips in 2018. It is the first general-purpose parallel (GPGPU) cloud computing chip and high-performance computing power in China. system provider. With the mission of “becoming an enabler of an intelligent society”, the company focuses on cloud server-level general-purpose parallel high-performance cloud computing chips, aiming at the data-driven technology market represented by cloud computing, artificial intelligence, and digital transformation to solve the core computing power Bottleneck problem. Focus on building self-controllable, world-class general-purpose, standard, high-performance cloud computing GPGPU chips to solve computing power problems from the chip side.

The Links:   LM64183PR LTM10C209A