What impact does the first AI chip of Pingtou people have on Ali?
On September 25th, Brother Pingtou took out the first 800-light AI chip. Ali, who loves fancy names, has not missed this opportunity-"Han Guang" is one of the three ancient Excalibur, echoing the embedded CPU Tie Xuan released two months ago and the system chip platform "No Sword" released 1 month ago, adding another member to the martial arts weapon camp.
According to Zhang Jianfeng, in the industry standard ResNet-50 test, the inferred performance of optical 800 reaches 78563 IPS, which is four times higher than that of the best AI chip in the industry. The energy efficiency ratio is 500 IPS/W, which is 3.3 times that of the second place. It is "the highest performance AI reasoning chip in the world".
Including light is not the only chip that entered the media field of vision in September this year. In the same month, Huawei released the AI chip Ascension 9 10, also known as "the strongest computing power in the world". There are two "strongest" in a month, but it is not the same thing. "The concept of the chip is as big as the Internet. There are as many chips as there are applications on the Internet. Many potato chips are cheaper than vegetables. " Zhang Jianfeng said.
But in fact, the similarity between Ascension 9 10 and containing light 800 is greater than the dissimilarity. All AI chips are deployed in the cloud, except that the former is a "training" chip and the latter is a "reasoning" chip, which belongs to two stages of deep learning. The training chip pays attention to absolute computing power, while the reasoning chip pays more attention to comprehensive indicators such as computing power per unit energy consumption, time delay and cost.
It is a new trend for cloud computing vendors to build their own cores in recent two years. Behind this is the rising cost of the whole industry. This month, Huawei Cloud &; Hou Jinlong, president of AI products and services, said that the server cost of Huawei's cloud data center has accounted for more than 60%, and it is difficult to make a profit from purchasing servers from outside in the future.
Intel, NVIDIA, AMD and other companies have long monopolized the chip manufacturing of cloud servers. According to DRAMeXchange, more than 90% of servers in the world are using Intel-led x86 architecture.
"Alibaba is a newcomer in the field of chips. Tie Xuan and Hanguang 800 are the first steps of Pingtou Long March, and we still have a long way to go. " Zhang Jianfeng said.
Huawei's thinking is different from Ali's. Although the chip is not sold separately, it is different from Ali's idea of "selling chips in the cloud". Huawei's approach is to sell chips in servers.
For Ali and Huawei, the development cost of self-developed chips is high and the cycle is long. Putting them into your own products first is a shortcut to ensure that research and development costs will not be wasted. At least two cloud factories can consume more than half of the chips on their own demand. In addition, Huawei's model has long-term significance for netting hardware manufacturers and building Huawei's ecology.