From @clausaasholm "the downgraded [Nvidia] H20 system that passes the embargo rules ... language model on a cluster of 2,048 Nvidia H800 GPUs and that it took two months, a total of 2.8 million ...

"The downgraded H20 system that passes the embargo rules for ... language model on a cluster of 2,048 Nvidia H800 GPUs and that it took two months, a total of 2.8 million GPU hours.

According to the paper, the company trained its V3 model on a cluster of 2,048 Nvidia H800 GPUs - crippled versions of the H100. The H800 launched in March 2023, to comply with US export ...

In a research paper released last month, DeepSeek said it had trained its V3 model using just 2,048 of Nvidia’s H800 chips. Nvidia specifically created the less-power H800 for sale to Chinese ...

Worse for Nvidia, the state-of-the-art V3 LLM was trained on just 2,048 of Nvidia’s H800 GPUs over two months, equivalent to about 2.8 million GPU hours, or about one-tenth the computing power ...

As for the hardware, DeepSeek used Nvidia H800 GPUs, which are modified from typically used H100 GPUs to abide by U.S. export restrictions. With a large number of consumers trying out DeepSeek's ...

Nvidia's tailored A800, H800, and H20 chips remain critical for AI growth, with 1 million H20 chips potentially generating $12B in sales. 3 Summer "Power Patterns" Are About to Trigger (One With ...

DeepSeek claimed its chatbot was trained on 2,000 Nvidia H800 GPUs at a cost of less than $6 million — though critics have cast doubt on that figure. DeepSeek's emergence roiled U.S. tech stocks ...