DeepSeek Reportedly betting on Huawei chips to train large models, with a new batch of chips expected as early as the 4th quarter
BIBIBI
AT A GLANCE
DeepSeek Liang Wenfeng said that adapting to domestic chips such as those made by Huawei had been designated as a key bet and must succeed.
Article
DeepSeek Betting on Huawei training chips, Liang Wenfeng says it "must succeed"
Beating AI news flash: The Information, citing people familiar with the matter, said DeepSeek CEO at the latest closed-door investor meeting, Liang Wenfeng identified training large models using domestic chips such as Huawei's as one of the company's biggest bets, saying it "must succeed."DeepSeek It expects to receive a new batch of Huawei chips usable for training as early as the fourth quarter of this year; the specific models and quantities have not been disclosed. Liang Wenfeng believes Huawei chips could catch up with Nvidia in just a few years and hopes DeepSeek to adapt to hardware beyond Nvidia's in advance.DeepSeek It is currently still using Nvidia chips to train models. Earlier this year, Liang Wenfeng May estimated that training the most advanced large models requires approximately 50000 chips Nvidia GB300, while using Huawei Ascend 950 would require approximately 200000 chips. At the time, the biggest issue was Huawei's insufficient supply.
Original link https://m.theblockbeats.info/flash/368334
Key points
01
The Information, citing people familiar with the matter, reported that Liang Wenfeng listed training large models using domestic chips such as those made by Huawei as one of the company's biggest bets at the latest closed-door investor meeting.
02
DeepSeek A new batch of Huawei chips suitable for training is expected to be obtained as early as the fourth quarter of this year.
03
The specific models and quantity of the Huawei chips have not been disclosed.
04
DeepSeek Models are still currently being trained using Nvidia chips.
05
Liang Wenfeng once estimated that training the most advanced large models requires approximately 50000 chips Nvidia GB300, while using Huawei Ascend 950 would require approximately 200000 chips.
AI-assisted interpretation
The following is analysis, separate from reported facts. Verify important claims independently.
This means DeepSeek is attempting to make model training compatible with chips other than Nvidia's, but is still currently using Nvidia chips; the availability of Huawei chips is also an issue that has been mentioned.
Why it matters to readers
If adaptation succeeds,DeepSeek there will be training hardware options beyond Nvidia; however, the materials do not state whether the adaptation has been completed.
Training chips affect the choice and quantity of hardware required for large-model development. The figures in the article show that different chips may require different quantities to complete comparable training.
Risks and unknowns
The information comes from people familiar with the matter cited by The Information; no original transcript of Liang Wenfeng's remarks was provided.
The specific models and quantities of Huawei chips have not yet been disclosed.
The information does not specify DeepSeek when the adaptation will be completed or whether the plan will succeed.
The specific year corresponding to “the fourth quarter of this year” is not explicitly stated in the original text.
Related Developments
Loading event timeline…
Related concepts
Large Language Models
This term is not in the glossary yet. Browse related concepts in the glossary.
