South Korean media: China's AI giant counters, DeepSeek R2 equipped with Huawei chips is about to be unveiled!
On May 4th, the South Korean media "Today Finance" published an article stating that before the release of the next-generation super-large language model "R2", Chinese AI startup company "DeepSeek" has attracted attention from the global AI industry.
It is understood that DeepSeek R2 is equipped with an advanced MoE architecture and has 1.2 trillion parameters, with text processing costs 97.3% cheaper than Open AI's GPT-4.
Notably, this model is fully trained using China's self-developed Huawei chipset (Ascend 910B), which is interpreted as a landmark case in reducing dependence on American Nvidia GPUs and strengthening the independence of China's AI industry.
The cluster based on Ascend 910B can provide 512 PetaFLOPS computing performance at FP16 precision, approximately 91% of the performance of an Nvidia A100 GPU cluster.
DeepSeek is an AI company co-founded by engineers from China's major technology companies in 2021. It gained international attention when it released the R1 model in January this year.
R1 demonstrated superior performance over competitors in multiple benchmark tests, thus making DeepSeek famous.
The R2 model is equipped with more than double the number of parameters compared to its predecessor and includes multimodal functions, expected to show scalability beyond simple conversational language models in the generative AI field.
Although DeepSeek R2 has not been officially released yet, it is expected to pose a threat to major AI companies such as OpenAI and Google's DeepMind in the international market due to its performance, cost competitiveness, and training capabilities based on domestic chipsets.
Original source: https://www.toutiao.com/article/1831189022393673/
Disclaimer: This article represents only the views of the author.