An embarrassing fact: once the Al chip is banned, the domestic AI large model will be GPL-free.
The hottest technology product in the first half of this year must be ChatGPT, an AI model, not only in foreign fire, Microsoft, Google, meta, Amazon, etc. all participated in the war, but also in China. Baidu, iFLYTEK, Ali, Huawei, etc. are also in the game. Although ChatGPT seems to have cooled down now, it is still called a thousand demon wars. And the popularity of this AI large model has also made NVIDIA a lot of money, becoming the world's first chip company with a market value of more than one trillion.
NVIDIA said that this is the iPhone era of AI, because in foreign countries everyone uses to train AI large models using NVIDIA's A100 and H100 chips. In China, because A100 and H100 are banned from sale in the United States, most of the special versions A800 and H800 launched by NVIDIA for the Chinese market are used, and these chips are currently difficult to find.
However, recently there is news that the United States intends to ban the sale of A800 and H800, not allowing NVIDIA to sell these powerful GPU chips to China. Nvidia Chief Financial Officer Colette Kress said that the United States' restrictions on the export of artificial intelligence chips to China will permanently lose the opportunity for the US industry, as a warning to the United States not to ban sales. But it's hard to predict whether there will be a ban on sales next. Industry insiders say the possibility of a lock-up is still very high.
So the question is, once NVIDIA A800 and H800 chips are also banned, how can China's AI models still play? Can domestic GPUs pick the beam? AI chips are GPUs, and the current GPU chips can be divided into three types, one is called AI for artificial intelligence - AI large models are also used, this is called FP - for supercomputing acceleration.
There is also a kind of rendering, which is ordinary game rendering, graphics applications. At present, domestic GPUs mainly have breakthroughs in rendering, such as Moore threaded GPUs, which are ordinary rendering GPUs.
AI large model training mainly needs to process high-granularity information, and requires double-precision floating-point computing capabilities, such as FP64 is a hard indicator of AI large model chips. At present, most of the domestic GPU chips do not have the ability required to support large model training, and mainly support single-precision floating-point computing FP32 computing requirements, and do not have the ability to process FP64.
Of course, in addition to this general AI large model, some vertical industry large models, some domestic GPUs can be replaced, but the performance is indeed not particularly strong.
So to tell the truth, if the United States really even A800, H800 and other chips are also banned, then the domestic AI general large model will indeed be available without cores in a short time, but the domestic GPU can not be replaced in a short time, but on some small vertical AI models, domestic GPUs can be replaced.
However, everyone also knows that as NVIDIA said, if the United States really wants to ban sales, then it will force the development of Chinese GPU chips, and the future will inevitably make gradual progress, and then let the GPU chips in the United States lose the opportunity forever, but it does take time, not in a short time.
