Once Claimed To Crush Nvidia!Wall -个 Technology: A Single Domestic AI Chip Is Not Strong But The Number Is Large

7News on the 10th, recently, Ding Yunfan, Vice President of Wall Science and Technology and Chief Architecture of AI Software, said when talking about calculating the bottleneck, saying that solving the computing power bottleneck needs to be considered from three dimensions: hardware cluster computing power, software effective computing power, differenceConstructing aggregate computing power.

He believes that even if the domestic AI chip has not strong computing power, it can improve computing power through comprehensive means to meet the needs of large domestic model training.

We 2020The first -generation product of the year was made by ChicletConstruction, foreign giants released this year's products such as Nvidia B100And Intel GAUDI 3Also adoptedIn the same way, they use the process, but they also need Chiclet to break through the Moore's law limit to enhance the computing power of the single card.Ding Yunfan said.

According to him, Wigmong Technology uses three -level asynchronous Checkpoint technology, combines the GPU's video memory and CPU memory, and even the backup system of multi -node memory to achieve a balance. At present, it can significantly reduce the cost of failure.

Previously, the first general GPU released the first general GPU of Gibi TechnologyChip BR100, Set a global computing power record, 16The floating point computing power reaches 1000TAbove, 8SetPoint computing power reaches more than 2000T, and the peak computing power of the single chip reaches the PFLOPS level.

At that time, the data given by Gyegle Technology showed that its first flagship product BR100 compared Nvidia's flagship GPU peak computing power on INT8, BF16, TF32/TF32+, and FP32.The performance advantage under FP32 data format has reached 13.1 times.

Statement on this site: This article is authorized by the author or related agencies. The purpose is to pass more information, which does not mean that this site agrees with its views, and this site does not guarantee or promise the authenticity of the content.If you need to reprint, please contact the column author. If the content of the article violates your rights and interests, please contact this site to delete in time.